JP5270025B2 - Parameter decoding apparatus and parameter decoding method - Google Patents

Abstract

Provided is a parameter decoding device for CELP encoded speech signals which performs parameter compensation process so as to suppress degradation of a main observation quality in a prediction quantization. The parameter decoding device includes amplifiers (305-1 to 305-M) which multiply inputted quantization prediction residual vectors x n-1 to x n-M by a weighting coefficient ² 1 to ² M . The amplifier (306) multiplies the preceding frame decoding LSF vector y n-1 by the weighting coefficient ² -1 . The amplifier (307) multiplies the code vector x n+1 outputted from a codebook (301) by the weighting coefficient ² 0 . An adder (308) calculates the total of the vectors outputted from the amplifiers (305-1 to 305-M), the amplifier (306), and the amplifier (307). A selector switch (309) selects the vector outputted from the adder (308) if the frame erasure coding B n of the current frame indicates that 'the n-th frame is an erased frame' and the frame erasure coding B n+1 of the next frame indicates that 'the n+1-th frame is a normal frame'.

Description

  The present invention relates to a parameter decoding apparatus and a parameter decoding method for decoding parameters encoded using a predictor.

  ITU-T Recommendation G. In speech codecs such as 729 and 3GPP AMR, some parameters obtained by analyzing speech signals are quantized by a predictive quantization method based on a moving average (MA) prediction model (Patent Document 1, Non-patent document 1, Non-patent document 2). The MA type predictive quantizer is a model that predicts the current parameter to be quantized by a linear sum of past quantized prediction residuals. In a code-excited linear prediction (CELP) type speech codec, It is used for prediction of line spectral frequency (LSF) parameters and energy parameters.

  Since the MA type predictive quantizer performs prediction with a weighted linear sum of quantized prediction residuals in the past finite number of frames, even if there is a transmission path error in the quantization information, the influence is limited to the number of finite frames. Limited. On the other hand, in an auto-regressive (AR) type predictive quantizer that uses past decoding parameters recursively, generally, a high prediction gain and quantization performance can be obtained, but the influence of errors extends for a long time. For this reason, the parameter quantizer for MA type prediction can realize higher error tolerance than the parameter quantizer for AR type prediction, and is used particularly for speech codecs for mobile communications.

  Here, conventionally, a parameter compensation method when a frame is lost on the decoding side has been studied. In general, compensation is performed using the parameters of the previous frame instead of the parameters of the lost frame. However, in the case of the LSF parameter, the parameter before the lost frame may be modified little by little by using a method of gradually approaching the average LSF or gradually decreasing in the case of the energy parameter.

  This method is usually used also in a quantizer using an MA type predictor, and in the case of an LSF parameter, a quantized prediction residual is generated so that a parameter generated in a compensation frame is decoded to generate an MA type. A process for updating the state of the predictor is performed (Non-Patent Document 1). In the case of energy parameters, an MA type predictor is used by using an average value of past quantized prediction residuals attenuated at a certain rate. The process of updating the state is performed (Patent Document 2, Non-Patent Document 1).

  There is also a method of interpolating the parameters of the lost frame after information on the return frame (normal frame) after the lost frame is obtained. For example, Patent Document 3 proposes a technique for regenerating the contents of an adaptive codebook by performing pitch gain interpolation.

JP-A-6-175695 JP-A-9-120297 JP 2002-328700 A

ITU-T Recommendation G. 729 3GPP TS 26.091

  The method of interpolating the parameters of the lost frame is used when predictive quantization is not performed, but when predictive quantization is performed, the encoded information is correctly received in the frame immediately after the lost frame. However, since the predictor is affected by the error in the previous frame and cannot obtain a correct decoding result, it is not generally used.

  As described above, the parameter quantization apparatus using the conventional MA type predictor does not perform the compensation process of the parameters of the lost frame by the interpolation method. May occur and may be a cause of deterioration of subjective quality.

  In addition, when predictive quantization is performed, a method of decoding parameters by simply interpolating the decoded quantized prediction residual may be considered, but even if the decoded quantized predictive residual fluctuates greatly, While the decoding parameter varies gradually between frames due to the weighted moving average, in this method, the decoding parameter also varies with the variation of the decoded quantization prediction residual. On the other hand, the deterioration of the subjective quality is increased.

  An object of the present invention has been made in view of such points, and in the case where predictive quantization is performed, a parameter decoding apparatus and a parameter that can perform parameter compensation processing so as to suppress deterioration of subjective quality It is to provide a decoding method.

  A parameter decoding apparatus according to an aspect of the present invention includes a prediction residual decoding unit that obtains a quantized prediction residual based on coding information included in a current frame to be decoded, and a parameter based on the quantized prediction residual. Parameter decoding means for decoding the current frame, the prediction residual decoding means, when the current frame is lost, the weighted linear sum of the parameter decoded in the past and the quantized prediction residual of the future frame A configuration for obtaining the quantized prediction residual of is taken.

  The parameter encoding apparatus according to an aspect of the present invention is obtained by analyzing an input signal to obtain an analysis parameter, predicting the analysis parameter using a prediction coefficient, and quantizing the prediction residual Coding means for obtaining a quantization parameter using a quantized prediction residual and the prediction coefficient, a plurality of sets of weighting coefficients are stored, the quantized prediction residual of the current frame, and the quantized prediction of two frames past For the residual and the quantization parameter in the past two frames, obtain a weighted sum using the set of weighting coefficients, and use the weighted sum to obtain a plurality of quantization parameters in the past one frame. Compensation means and a plurality of the quantization parameters in the past of one frame obtained by the previous frame compensation means are obtained by the analysis means in the past of one frame. A determination to select one of the quantization parameters in the past of one frame and to select and encode a weighting coefficient set corresponding to the selected quantization parameter in the past of the one frame compared with the analysis parameter; The structure which comprises a means is taken.

  The parameter decoding method according to an aspect of the present invention includes a prediction residual decoding step for obtaining a quantized prediction residual based on coding information included in a current frame to be decoded, and the quantized prediction residual. A parameter decoding step for decoding parameters, and in the prediction residual decoding step, when the current frame is lost, a weighted linear sum of a parameter decoded in the past and a quantized prediction residual of a future frame is obtained. A method for obtaining the quantized prediction residual of the current frame is adopted.

  According to the present invention, when predictive quantization is performed, if the current frame is lost, the weights of previously decoded parameters, quantized prediction residuals of past frames, and quantized prediction residuals of future frames are weighted. By obtaining the quantized prediction residual of the current frame by a linear sum, it is possible to perform parameter compensation processing so as to suppress deterioration in subjective quality.

The block diagram which shows the main structures of the speech decoding apparatus which concerns on Embodiment 1 of this invention. The figure which shows the internal structure of the LPC decoding part of the audio | voice decoding apparatus which concerns on Embodiment 1 of this invention. The figure which shows the internal structure of the code vector decoding part in FIG. The figure which shows an example of the result of having performed normal processing when there is no lost frame The figure which shows an example of the result of having performed the compensation process of this Embodiment The figure which shows an example of the result of having performed the conventional compensation processing The figure which shows an example of the result of having performed the conventional compensation processing The block diagram which shows the main structures of the speech decoding apparatus which concerns on Embodiment 2 of this invention. The block diagram which shows the internal structure of the LPC decoding part in FIG. The block diagram which shows the internal structure of the code vector decoding part in FIG. The block diagram which shows the main structures of the speech decoding apparatus which concerns on Embodiment 3 of this invention. The block diagram which shows the internal structure of the LPC decoding part in FIG. The block diagram which shows the internal structure of the code vector decoding part in FIG. The block diagram which shows the internal structure of the gain decoding part in FIG. The block diagram which shows the internal structure of the prediction residual decoding part in FIG. The block diagram which shows the internal structure of the sub-frame quantization prediction residual generation part in FIG. Block diagram showing the main configuration of a speech encoding apparatus according to Embodiment 5 of the present invention. The block diagram which shows the structure of the audio | voice signal transmission apparatus and audio | voice signal receiving apparatus which comprise the audio | voice signal transmission system which concerns on Embodiment 6 of this invention. The figure which shows the internal structure of the LPC decoding part of the speech decoding apparatus which concerns on Embodiment 7 of this invention. The figure which shows the internal structure of the code vector decoding part in FIG. The block diagram which shows the main structures of the speech decoder based on Embodiment 8 of this invention. The figure which shows the internal structure of the LPC decoding part of the speech decoding apparatus which concerns on Embodiment 8 of this invention. The figure which shows the internal structure of the code vector decoding part in FIG. The figure which shows the internal structure of the LPC decoding part of the speech decoding apparatus which concerns on Embodiment 9 of this invention. The figure which shows the internal structure of the code vector decoding part in FIG. Block diagram showing the main configuration of a speech decoding apparatus according to Embodiment 10 of the present invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. In each of the following embodiments, a case where the parameter decoding apparatus / parameter encoding apparatus of the present invention is applied to a CELP speech decoding apparatus / speech encoding apparatus will be described as an example.

(Embodiment 1)
FIG. 1 is a block diagram showing the main configuration of the speech decoding apparatus according to Embodiment 1 of the present invention. In the speech decoding apparatus 100 shown in FIG. 1, encoded information transmitted from an encoding apparatus (not shown) is transmitted by a demultiplexing unit 101 to a fixed codebook code F _{n + 1} , adaptive codebook code A _{n + 1} , gain code G _{n + 1} , and LPC (Linear Prediction Coefficients) code Ln _{+ 1} . Separately, frame erasure code B _{n + 1} is input to speech decoding apparatus 100. Here, the subscript n of each code represents a frame number to be decoded. That is, in FIG. 1, the encoded information in the (n + 1) th frame (hereinafter referred to as “next frame”) following the nth frame to be decoded (hereinafter referred to as “current frame”) is separated.

The fixed codebook code F _{n + 1} is supplied to the fixed codebook vector (FCV) decoding unit 102 and the adaptive codebook code _{An + 1} is supplied to the adaptive codebook vector (Adaptive Codebook Vector (ACV)) decoding unit 103. G _{n + 1} is input to the gain decoding unit 104, and the LPC code L _{n + 1} is input to the LPC decoding unit 105. The frame erasure code B _{n + 1} is input to all of the FCV decoding unit 102, the ACV decoding unit 103, the gain decoding unit 104, and the LPC decoding unit 105.

When the frame erasure code B _n indicates that “the nth frame is a normal frame”, the FCV decoding unit 102 generates a fixed codebook vector using the fixed codebook code F _n and generates a frame erasure code. When B _n indicates that “the nth frame is a lost frame”, a fixed codebook vector is generated by frame loss compensation (concealment) processing. The generated fixed codebook vector is input to gain decoding section 104 and amplifier 106.

ACV decoding unit 103, when the frame erasure code B _n indicates that "the n-th frame is normal frame" generates an adaptive codebook vector using the adaptive codebook code A _n, frame erasure code When B _n indicates that “the nth frame is a lost frame”, an adaptive codebook vector is generated by frame loss compensation (concealment) processing. The generated adaptive codebook vector is input to the amplifier 107.

When the frame erasure code B _n indicates that “the nth frame is a normal frame”, the gain decoding unit 104 uses the gain code G _n and the fixed codebook vector to determine the fixed codebook gain and the adaptive codebook. A gain is generated, and when the frame erasure code B _n indicates that “the nth frame is a erasure frame”, a fixed codebook gain and an adaptive codebook gain are generated by frame erasure compensation (concealment) processing. . The generated fixed codebook gain is input to the amplifier 106, and the generated adaptive codebook gain is input to the amplifier 107.

When the frame erasure code B _n indicates that “the nth frame is a normal frame”, the LPC decoding unit 105 decodes the LPC parameter using the LPC code L _n , and the frame erasure code B _n is “ If it is indicated that “the nth frame is a lost frame”, the LPC parameter is decoded by frame loss compensation (concealment) processing. The decrypted LPC parameter is input to the LPC synthesis unit 109. Details of the LPC decoding unit 105 will be described later.

  Amplifier 106 multiplies the fixed codebook gain output from gain decoding section 104 by the fixed codebook vector output from FCV decoding section 102, and outputs the multiplication result to adder 108. Amplifier 107 multiplies the adaptive codebook gain output from gain decoding section 104 by the adaptive codebook vector output from ACV decoding section 103, and outputs the multiplication result to adder 108. The adder 108 adds the fixed codebook vector after multiplication of the fixed codebook gain output from the amplifier 106 and the adaptive codebook vector after multiplication of the adaptive codebook gain output from the amplifier 107, and adds the result (hereinafter, referred to as "addition code"). “Sum vector”) is output to the LPC synthesis unit 109.

  The LPC synthesis unit 109 configures a linear prediction synthesis filter using the decoded LPC parameters output from the LPC decoding unit 105, drives the linear prediction synthesis filter using the sum vector output from the adder 108 as a drive signal, and drives The combined signal obtained as a result of is output to the post filter 110. The post filter 110 performs formant emphasis and pitch emphasis processing on the synthesized signal output from the LPC synthesizing unit 109, and outputs the result as a decoded speech signal.

  Next, details of the parameter compensation processing according to the present embodiment will be described using an example in which LPC parameters are compensated. FIG. 2 is a diagram showing an internal configuration of the LPC decoding unit 105 in FIG.

The LPC code L _{n + 1} is input to the buffer 201 and the code vector decoding unit 203, and the frame erasure code B _{n + 1} is input to the buffer 202, the code vector decoding unit 203, and the selector 209.

The buffer 201 holds the LPC code L _{n + 1} of the next frame for one frame and outputs it to the code vector decoding unit 203. LPC code output from the buffer 201 to the code vector decoding unit 203, a result which is held for one frame in buffer 201, the LPC code _{L n} of the current frame.

The buffer 202 holds the frame erasure code B _{n + 1} of the next frame for one frame and outputs it to the code vector decoding unit 203. The frame erasure code output from the buffer 202 to the code vector decoding unit 203 becomes the frame erasure code _Bn of the current frame as a result of being held in the buffer 202 for one frame.

The code vector decoding unit 203 includes quantized prediction residual vectors x _{n-1 to} x _nM of the past M frames, the decoded LSF vector y _{n-1 of} the previous frame, the LPC code L _{n + 1 of} the next frame, and the frame erasure of the next frame The code B _{n + 1} , the LPC code L _n of the current frame and the frame erasure code B _n of the current frame are input, and a quantized prediction residual vector x _n of the current frame is generated based on these information, and the buffer 204-1 and Output to amplifier 205-1. Details of the code vector decoding unit 203 will be described later.

The buffer 204-1 holds the quantized prediction residual vector _xn of the current frame for one frame, and outputs it to the code vector decoding unit 203, the buffer 204-2, and the amplifier 205-2. The quantized prediction residual vector input to these becomes the quantized prediction residual vector x _{n-1 of the} previous frame as a result of being held for one frame in the buffer 204-1. Similarly, the buffer 204-i (i is 2 to M-1) holds the quantized prediction residual vector x _{n-j + 1} for one frame, respectively, and the code vector decoding unit 203 and the buffer 204- ( i + 1) and the amplifier 205- (i + 1). The buffer 204-M holds the quantized prediction residual vector x _{n-M + 1} for one frame and outputs it to the code vector decoding unit 203 and the amplifier 205- (M + 1).

The amplifier 205-1 multiplies the quantized prediction residual vector x _n by a predetermined MA prediction coefficient α ₀ and outputs the result to the adder 206. Similarly, the amplifier 205-j (j is 2 to M + 1) multiplies the quantized prediction residual vector x _{n-j + 1} by a predetermined MA prediction coefficient α _j−1 and outputs the result to the adder 206. Note that the MA prediction coefficient set may be one type of fixed value. In 729, two types of sets are prepared, and which set is used for decoding is determined on the encoder side, and is encoded and transmitted as part of the information of the LPC code Ln. In this case, the LPC decoding unit 105 includes a set of MA prediction coefficients as a table, and uses a set designated on the encoder side as α _{0 to} α _{M in} FIG.

The adder 206 calculates the sum of the quantized prediction residual vectors after multiplication of the MA prediction coefficients output from the amplifiers 205-1 to 205- (M + 1), and stores the decoded LSF vector y _n as the calculation result in the buffer 207. And output to the LPC converter 208.

The buffer 207 holds the decoded LSF vector y _n for one frame and outputs it to the code vector decoding unit 203. As a result, the decoded LSF vector output from the buffer 207 to the code vector decoding unit 203 is the decoded LSF vector y _{n-1 of the} previous frame.

LPC converting section 208 converts the decoded LSF vector y _n in the linear prediction coefficient (decoded LPC parameter), and outputs to the selector 209.

The selector 209 selects the decoded LPC parameter output from the LPC conversion unit 208 based on the frame erasure code B _n of the current frame and the frame erasure code B _{n + 1} of the next frame or the decoded LPC parameter in the previous frame output from the buffer 210. Choose one. Specifically, when the frame erasure code _Bn of the current frame indicates that “the nth frame is a normal frame”, or when the frame erasure code _{Bn + 1 of} the next frame is “the n + 1th frame is a normal frame” In the case where “there is”, the decoded LPC parameter output from the LPC conversion unit 208 is selected, and the frame erasure code B _n of the current frame indicates that “the nth frame is a lost frame” When the frame erasure code B _{n + 1 of} the next frame indicates that “the (n + 1) th frame is a erasure frame”, the decoding LPC parameter in the previous frame output from the buffer 210 is selected. Then, the selector 209 outputs the selection result to the LPC synthesis unit 109 and the buffer 210 as a final decoded LPC parameter. Note that when the selector 209 selects the decoded LPC parameter in the previous frame output from the buffer 210, it is not actually necessary to perform all the processing from the code vector decoding unit 203 to the LPC conversion unit 208, and the buffer 204-1. Only the process of updating the content of ˜204-M needs to be performed.

  The buffer 210 holds the decoded LPC parameter output from the selector 209 for one frame and outputs it to the selector 209. As a result, the decoded LPC parameter output from the buffer 210 to the selector 209 is the decoded LPC parameter of the previous frame.

  Next, the internal configuration of the code vector decoding unit 203 in FIG. 2 will be described in detail with reference to the block diagram of FIG.

The code book 301 generates a code vector specified by the LPC code L _n of the current frame and outputs the code vector to the changeover switch 309, and generates a code vector specified by the LPC code L _{n + 1 of} the next frame and supplies it to the amplifier 307. Output. As already mentioned, ITU-T Recommendation G. Information identifying the MA prediction coefficient set to 729 in LPC code L _n is also included, in this case, although the LPC code L _n is also used for decoding the addition to the MA prediction coefficients for decoding the code vector, wherein the Description is omitted. The code book may have a multi-stage configuration or a split configuration. For example, ITU-T Recommendation G. 729 is a two-stage configuration, and the second stage is a codebook configuration divided into two (split). Also, vectors output from multi-stage or divided codebooks are not usually used as they are, and when the interval between orders is extremely narrow or the order is reversed, the minimum interval becomes a specific value. In general, a process for guaranteeing the order and maintaining the order is performed.

The quantized prediction residual vectors x _{n-1 to} x _nM of the past M frames are input to the corresponding amplifiers 302-1 to 302-M and the corresponding amplifiers 305-1 to 305-M, respectively.

The amplifiers 302-1 to 302-M multiply the input quantized prediction residual vectors x _{n−1 to} x _nM by MA prediction coefficients α _{1 to} α _M and output the result to the adder 303. As described above, the ITU-T recommendation G.I. In the case of 729, there are two sets of MA prediction coefficients, and information on which one to use is included in the LPC code L _n . Further, since the LPC code L _n is lost in the lost frame in which these multiplications are performed, the MA prediction coefficient set used in the previous frame is actually used. That is, the MA prediction coefficient set information decoded from the LPC code L _{n−1 of} the previous frame is used. If the previous frame is also a lost frame, the information of the previous frame is further used.

The adder 303 calculates the sum of the respective quantized prediction residual vectors after multiplication of the MA prediction coefficients output from the amplifiers 302-1 to 302-M, and outputs a vector as a calculation result to the adder 304. The adder 304 subtracts the vector output from the adder 303 from the decoded LSF vector y _{n−1 of} the previous frame output from the buffer 207, and outputs a vector as a calculation result to the changeover switch 309.

The vector output from the adder 303 is a predicted LSF vector predicted by the MA type predictor in the current frame, and the adder 304 calculates the quantized prediction in the current frame necessary for generating the decoded LSF vector of the previous frame. Processing for obtaining a residual vector is performed. That is, amplifiers 302-1 to 302-M, adder 303, and adder 304 calculate vectors so that decoded LSF vector y _{n-1 of} the previous frame becomes decoded LSF vector y _{n of the} current frame.

The amplifiers 305-1 to 305-M multiply the input quantized prediction residual vectors x _{n-1 to} x _nM by weighting coefficients β _{1 to} β _M and output the result to the adder 308. Amplifier 306 multiplies the weighting factor beta _-1 previously decoded LSF vector y _n-1 of the frame output from the buffer 207, and outputs to the adder 308. The amplifier 307 multiplies the code vector x _{n + 1} output from the code book 301 by the weighting coefficient β ₀ and outputs the result to the adder 308.

The adder 308 calculates the sum of the vectors output from the amplifiers 305-1 to 305-M, the amplifier 306, and the amplifier 307, and outputs the code vector that is the calculation result to the changeover switch 309. That is, the adder 308 calculates a vector by weighted addition of the code vector specified by the LPC code L _{n + 1} of the next frame, the decoded LSF vector of the previous frame, and the quantized prediction residual vector of the past M frames. Yes.

When the frame erasure code _Bn of the current frame indicates that “the nth frame is a normal frame”, the changeover switch 309 selects the code vector output from the codebook 301 and quantizes the code frame of the current frame. The prediction residual vector x _n is output. On the other hand, when the switch 309 indicates that the frame erasure code B _n of the current frame indicates “the nth frame is a erasure frame”, which information the frame erasure code B _{n + 1} of the next frame has. To further select a vector to be output.

That is, when the frame erasure code B _{n + 1 of} the next frame indicates that “the (n + 1) th frame is a erasure frame”, the changeover switch 309 selects the vector output from the adder 304 and replaces it with the current frame. The quantized prediction residual vector x _n is output. In this case, it is not necessary to perform the process of generating a vector from the code book 301 and the amplifiers 305-1 to 305-M to the adder 308.

Also, when the frame erasure code B _{n + 1 of} the next frame indicates that “the (n + 1) th frame is a normal frame”, the changeover switch 309 selects the vector output from the adder 308, and this is selected as the current frame. The quantized prediction residual vector x _n is output. In this case, it is not necessary to perform a process of generating a vector from the amplifiers 302-1 to 302-M to the adder 304.

  As described above, according to the present embodiment, when the current frame is lost, if the next frame is normally received, the parameters decoded in the past and the quantized prediction residual of the frame received in the past are stored. Compensated quantized prediction by performing compensation processing of the decoded quantized prediction residual of the LSF parameter of the current frame by weighted addition processing (weighted linear sum) dedicated to compensation processing using the difference and the quantized prediction residual of the future frame The LSF parameter is decoded using the residual. As a result, it is possible to realize higher compensation performance than repeatedly using past decoded LSF parameters.

  Hereinafter, the result of performing the compensation processing of the present embodiment will be described with reference to specific examples using FIGS. 4 to 7 in comparison with the prior art. In FIGS. 4 to 7, ◯ indicates a decoded quantized prediction residual, ● indicates a decoded quantized prediction residual obtained by compensation processing, ◇ indicates a decoding parameter, and ♦ indicates decoding obtained by the compensation processing. Each parameter is shown.

Figure 4 is a diagram showing an example of the results when no erased frame by usual processing to obtain the decoding parameter y _n of the n-th frame by the following equation (1) from decoded quantized prediction residual Is. In the equation (1), c _n is the decoded quantized prediction residual of the n-th frame.
_{y n = 0.6c n + 0.3c n} -1 + 0.1c n-2 ··· (1)

  FIG. 5 is a diagram illustrating an example of a result of performing the compensation processing of the present embodiment, and FIGS. 6 and 7 are diagrams illustrating an example of a result of performing the conventional compensation processing. 5, 6, and 7, it is assumed that the nth frame is lost and the other frames are normal frames.

Compensation processing of this embodiment shown in FIG. 5, as the variation between frames of the decoded parameter becomes gentle, distance decoding parameter y _n of the decoded parameter y _n-1 of the (n-1) th frame n-th frame , And the sum D of the distances between the decoding parameter y _n of the nth frame and the decoding parameter y _{n + 1} of the ( _{n + 1) th} frame is as follows so that D is defined by the following equation (2): obtaining dequantized prediction residual c _n of the n-th frame which has lost using equation (3).

In the compensation processing of this embodiment, the decoded parameter y _n of the lost nth frame is obtained by the above equation (1) using the decoded quantization prediction residual c _n obtained by the equation (3). . As a result, as is clear from the comparison between FIG. 4 and FIG. 5, the decoding parameter y _n obtained by the compensation processing of the present embodiment is almost the same as that obtained by normal processing when there is no lost frame. Value.

In contrast, conventional compensation process shown in FIG. 6, when the n-th frame is lost, using the decoded parameter y _n-1 of the (n-1) th frame as it is decoded parameter y _n of the n-th frame. In the conventional compensation process shown in FIG. 6, the decoded quantization prediction residual c _n of the nth frame is obtained by the inverse calculation of the above equation (1).

In this case, since the variation of the decoding parameter due to the variation of the decoded quantization prediction residual is not taken into account, as is clear from the comparison between FIG. 4 and FIG. 6, the decoding parameter obtained by the conventional compensation processing of FIG. y _n is the value that obtained by conventional processing when there is no erased frame largely differs. Further, since the decoded quantization prediction residual c _n of the nth frame is also different, the decoding parameter y _{n + 1 of the (n + 1) th} frame obtained by the conventional compensation process of FIG. 6 is also a normal process when there is no lost frame. The value is different from that obtained by.

Further, the conventional compensation process shown in FIG. 7 is to obtain the decoded quantized prediction residual by interpolation, and when the nth frame disappears, the decoded quantized predicted residual c _{n of} the (n−1) th frame. _-1 and using the decoded quantized average value of the prediction residuals c _{n + 1} of the (n + 1) th frame as a decoded quantized prediction residual c _n of the n-th frame.

The conventional compensation process shown in FIG. 7, using the decoded quantization prediction residual c _n obtained by the interpolation, by the above formula (1), obtains a decoded parameter y _n of the n-th frame which has lost .

As a result, as is apparent from a comparison of FIG. 4 and FIG. 7, decoding parameter y _n obtained by conventional compensation process of Figure 7, is the value that obtained by conventional processing when there is no lost frame It will be very different. This is because, even if the decoded quantization prediction residual fluctuates greatly, the decoding parameter gradually fluctuates between frames due to the weighted moving average. This is because the decoding parameters also fluctuate. Since the decoded quantization prediction residual c _n of the nth frame is also different, the decoding parameter y _{n + 1 of the (n + 1) th} frame obtained by the conventional compensation process of FIG. 7 is also a normal process when there is no lost frame. The value is different from that obtained by.

(Embodiment 2)
FIG. 8 is a block diagram showing the main configuration of the speech decoding apparatus according to Embodiment 2 of the present invention. The speech decoding apparatus 100 shown in FIG. 8 differs from FIG. 1 only in that compensation mode information En _{+ 1} is further added as a parameter input to the LPC decoding unit 105.

FIG. 9 is a block diagram showing an internal configuration of the LPC decoding unit 105 in FIG. The LPC decoding unit 105 shown in FIG. 9 differs from FIG. 2 only in that compensation mode information En _{+ 1} is further added as a parameter input to the code vector decoding unit 203.

  FIG. 10 is a block diagram showing an internal configuration of code vector decoding section 203 in FIG. The code vector decoding unit 203 shown in FIG. 10 differs from FIG. 3 only in that a coefficient decoding unit 401 is further added.

The coefficient decoding unit 401 stores a plurality of sets of weighting coefficients (β _{−1 to} β _M ) (hereinafter referred to as “coefficient sets”), and selects _one of the coefficient sets according to the input compensation mode En _{+ 1} . A set of weighting factors is selected and output to amplifiers 305-1 to 305-M, 306, and 307.

  As described above, according to the present embodiment, in addition to the features described in the first embodiment, a plurality of weighted addition weighting coefficient sets for performing compensation processing are prepared, and which weighting coefficient set is set on the encoder side. After confirming whether high compensation performance can be obtained by using, information for specifying the optimum set is transmitted to the decoder side, and the decoder side sets the specified weighting coefficient set based on the received information. Since the compensation processing is performed using this, higher compensation performance than that of the first embodiment can be obtained.

(Embodiment 3)
FIG. 11 is a block diagram showing the main configuration of the speech decoding apparatus according to Embodiment 3 of the present invention. Compared with FIG. 8, speech decoding apparatus 100 shown in FIG. 11 further includes a separation unit 501 that separates LPC code L _{n + 1} input to LPC decoding unit 105 into two types of codes V _{n + 1} and K _{n + 1.} The only difference is that A code V is a code for generating a code vector, and a code K is an MA prediction coefficient code.

FIG. 12 is a block diagram showing the internal configuration of the LPC decoding unit 105 in FIG. The codes V _n and V _{n + 1} that generate the code vector are used in the same manner as the LPC codes L _n and L _{n + 1,} and thus the description thereof is omitted. Compared to FIG. 9, the LPC decoding unit 105 illustrated in FIG. 12 further includes a buffer 601 and a coefficient decoding unit 602, and further includes an MA prediction coefficient code K _{n + 1} as a parameter input to the code vector decoding unit 203. Only the difference is.

The buffer 601 holds the MA prediction coefficient code K _{n + 1} for one frame and outputs it to the coefficient decoding unit 602. As a result, MA prediction coefficient code output from the buffer 601 to the coefficient decoding section 602, a preceding frame MA predictive coefficient code _{K n.}

Coefficient decoding section 602 stores a plurality of types of coefficient sets, specifies coefficient sets by frame erasure codes B _n , B _{n + 1} , compensation mode E _{n + 1} and MA prediction coefficient code K _n , and amplifiers 205-1 to 205- (M + 1) ). Here, the coefficient decoding unit 602 specifies the coefficient set in the following three ways.

If the frame erasure code B _n input indicates that "the n-th frame is normal frame", the coefficient decoding section 602 selects the coefficient set specified by MA prediction coefficient code K _n.

When the input frame erasure code _Bn indicates that “the nth frame is a erasure frame” and the frame erasure code _{Bn + 1} indicates that “the n + 1th frame is a normal frame”, the coefficient decoding unit 602 The coefficient set to be selected is determined using the compensation mode E _{n + 1} received as the parameter of the (n + 1) th frame. For example, if compensation mode code E _{n + 1} is determined in advance to show the mode of MA predictive coefficients to be used in the n-th frame is compensated frame, the compensation mode code E _{n + 1} as an alternative to the MA prediction coefficient code K _n as Can be used.

Further, when the input frame erasure code _Bn indicates that “the nth frame is a erasure frame” and the frame erasure code _{Bn + 1} indicates that “the n + 1th frame is an erasure frame”, information that can be used Is only information on the coefficient set used in the previous frame, the coefficient decoding unit 602 repeatedly uses the coefficient set used in the previous frame. Alternatively, a predetermined coefficient set may be used in a fixed manner.

FIG. 13 is a block diagram showing the internal configuration of the code vector decoding unit 203 in FIG. The code vector decoding unit 203 illustrated in FIG. 13 is different from that illustrated in FIG. 10 in that the coefficient decoding unit 401 selects a coefficient set using both the compensation mode E _{n + 1} and the MA prediction coefficient code K _{n + 1} .

  In FIG. 13, the coefficient decoding unit 401 includes a plurality of weighting coefficient sets, and the weighting coefficient sets are prepared according to the MA prediction coefficients used in the next frame. For example, when there are two types of MA prediction coefficient sets and one is mode 0 and the other is mode 1, the following sets of dedicated weighting coefficient sets and the next set of MA prediction coefficients for the next frame are mode 0. This is composed of a dedicated weighting coefficient set group when the MA prediction coefficient set of the frame is mode 1.

In this case, the coefficient decoding unit 401 determines one of the weighting coefficient set groups based on the MA prediction coefficient code K _{n + 1} , and selects one weighting coefficient set from among the coefficient sets according to the input compensation mode E _{n + 1.} Select and output to amplifiers 305-1 to 305-M, 306, 307.

Hereinafter, an example of how to determine the weighting coefficients β _{−1 to} β _M will be described. As already described, when the nth frame is lost and the (n + 1) th frame is received, the final decoding parameters are unknown in both frames even if the quantized prediction residual in the (n + 1) th frame can be correctly decoded. . For this reason, unless some assumption (constraint condition) is set, the decoding parameters of both frames are not uniquely determined. Therefore, the decoding parameter of the nth frame and the decoding parameter of the n−1th frame are set so that the decoding parameters of the nth frame and the (n + 1) th frame are as far as possible from the decoding parameters of the (n−1) th frame that has already been decoded. The quantization prediction residual y _n is _expressed by the following equation (4) so that D ^(j) which is the sum of the distance and the distance between the decoding parameter in the (n + 1) th frame and the decoding parameter in the nth frame is minimized. Ask.

When the parameter is LSF, x _n ^(j) , y _n ^(j) , α _i ^(j) , α ′ _i ^(j) in equation (4) are as follows.
x _n ^(j) : quantized prediction residual y _n ^(j) of the LSF parameter in the nth frame: ^j j component α _i ^(j) of the decoded LSF parameter in the nth frame: MA in the nth frame J-th component α ′ _i ^(j) of the i-th component of the prediction coefficient set: j-th component of the i-th component of the MA prediction coefficient set in the ⁽ n + 1 ⁾ th frame M: MA prediction order

Here, when the equation obtained by partial differentiation of D ^(j) with x _n ^(j) and setting it to 0 is solved with respect to x _n ^(j) , x _n ^(j) is expressed by the following equation (5). It is represented by

In equation (5), β _i ^(j) is a weighting coefficient and is expressed by α _i ^(j) and α ′ _i ^(j) . That is, when there is only one set of MA prediction coefficients, there is also only one set of weighting coefficients β _i ^(j) , but when there are multiple MA prediction coefficient sets, α _i ^(j) and α ′ A combination of a plurality of types of weighting coefficients can be obtained by combining _i ^(j) .

  For example, ITU-T Recommendation G. In the case of 729, since there are two types of MA prediction coefficient sets, if these are set to mode 0 and mode 1, both the n-th frame and the n + 1-th frame are mode 0, and the n-th frame is mode 0. Four types of sets are conceivable: when the n + 1-th frame is mode 1, the n-th frame is mode 1, the n + 1-th frame is mode 0, and both the n-th and n + 1-th frames are mode 1. There are several possible ways of determining which of the four types of sets to use.

  The first method generates an nth frame decoded LSF and an n + 1th frame decoded LSF on the encoder side using all four types of sets, and analyzes the generated nth frame decoded LSF and the input signal. The Euclidean distance between the obtained unquantized LSF and the decoded LSF of the generated (n + 1) th frame and the unquantized LSF obtained by analyzing the input signal are calculated. This is a method of selecting one set of weighting coefficients β that minimizes the sum of the Euclidean distances, encoding the selected set with 2 bits, and transmitting it to the decoder. In this case, ITU-T Recommendation G. In addition to the encoding information of 729, 2 bits per frame are necessary for encoding the coefficient set β. Note that ITU-T Recommendation G. Adopting a weighted Euclidean distance, as used in the 729 LSF quantization, can result in even better audio quality.

The second method uses the MA prediction coefficient mode information of the (n + 1) th frame and sets the number of additional bits per frame to 1 bit. Since the decoder knows the mode information of the MA prediction coefficient of the (n + 1) th frame, the combinations of α _i ^(j) and α ′ _i ^(j) are limited to two. That is, when the MA prediction mode of the (n + 1) th frame is mode 0, the combination of the MA prediction modes of the nth frame and the (n + 1) th frame is either (0-0) or (1-0). The set of β can be limited to two types. On the encoder side, using the set of these two types of weighting coefficients β, in the same way as in the first method, the one with the smaller error from the unquantized LSF is selected, encoded, and transmitted to the decoder. Good.

  The third method is a method in which selection information is not transmitted at all. The set of weighting coefficients to be used is only two types of combinations of MA prediction modes (0-0) or (1-1), and the (n + 1) th frame. This is a method of selecting the former when the mode of the MA prediction coefficient at 0 is 0, and selecting the latter when the mode is 1. Alternatively, a method of fixing the lost frame mode to a specific mode, such as (0-0) or (0-1), may be used.

  In addition, in a frame in which the input signal can be determined to be stationary, a method of making the decoding parameters of the (n−1) th frame and the nth frame equal to each other as in the conventional method, or the (n + 1) th frame and the nth frame. A method using a set of weighting factors β obtained under the assumption that the decoding parameters of the frames are equal is also conceivable.

  For determination of continuity, pitch period information of the (n−1) th frame and (n + 1) th frame, mode information of MA prediction coefficient, and the like can be used. That is, when the difference in pitch period decoded between the (n-1) th frame and the (n + 1) th frame is small, the method is determined to be stationary, or the mode information of the MA prediction coefficient decoded at the (n + 1) th frame is a stationary frame. If a mode suitable for encoding (that is, a mode in which a high-order MA prediction coefficient also has a certain amount of weight) is selected, a method of determining that it is stationary can be considered.

  As described above, in this embodiment, in addition to the second embodiment, there are two types of MA prediction coefficient modes. Therefore, different sets of MA prediction coefficients can be used in the stationary section and the other sections. The performance of the LSF quantizer can be further improved.

  In addition, by using the weighting coefficient set of Expression (5) that minimizes Expression (4), the decoded LSF parameter in the normal frame that is the frame subsequent to the lost frame and the lost frame becomes the LSF of the previous frame of the lost frame. It is guaranteed that the value does not deviate significantly from the parameter. For this reason, even when the decoded LSF parameter of the next frame is unknown, the risk of compensating in the wrong direction while effectively using the reception information (quantized prediction residual) of the next frame, that is, The risk of greatly deviating from the correct decryption LSF parameter can be minimized.

  Further, if the second method is used as the compensation mode selection method, the MA prediction coefficient mode information can be used as a part of the information for specifying the weighting coefficient set for the compensation process, so that additional transmission is performed. Information on the weighting coefficient set for compensation processing can also be reduced.

(Embodiment 4)
FIG. 14 is a block diagram showing the internal configuration of gain decoding section 104 in FIG. 1 (the same applies to gain decoding section 104 in FIGS. 8 and 11). In the present embodiment, ITU-T Recommendation G. Similarly to the case of G.729, gain decoding is performed once in each subframe, and one frame is composed of two subframes. In FIG. 14, n is a frame number and m is a subframe number (the nth frame in the nth frame). The subframe numbers of the 1st subframe and the 2nd subframe are m and m + 1), and the gain codes ( _Gm , _{Gm + 1} ) for the 2nd subframe of the nth frame are sequentially decoded.

In FIG. 14, gain code G _{n + 1 of the (n + 1) th} frame is input to gain decoding section 104 from demultiplexing section 101. The gain code G _{n + 1} is input to the separation unit 700 and separated into the gain code G _{m + 2} of the first subframe of the (n + 1) th frame and the gain code G _{m + 3} of the second subframe. The demultiplexing into the gain codes G _{m + 2} and G _{m + 3} may be performed by the demultiplexing unit 101.

Gain decoding section 104 sequentially decodes the decoding gain of subframe m and the decoding gain of subframe m + 1 using G _m , G _{m + 1} , G _{m + 2} and G _{m + 3} generated from input G _n and G _{n + 1.} .

Hereinafter, in FIG. 14 will be described the operation of each part of the gain decoding unit 104 in decoding the gain code G _m.

The gain code G _{m + 2} is input to the buffer 701 and the prediction residual decoding unit 704, and the frame erasure code B _{n + 1} is input to the buffer 703, the prediction residual decoding unit 704, and the selector 713.

Since the buffer 701 holds the input gain code for one frame and outputs it to the prediction residual decoding unit 704, the gain code output to the prediction residual decoding unit 704 is the gain code of one frame before. That is, if the gain code input to the buffer 701 is _{G m + 2,} the gain code output is _{G m.} The buffer 702 performs the same processing as that of 701. That is, the input gain code is held for one frame and output to the prediction residual decoding unit 704. The only difference is that the input / output of the buffer 701 is the gain code of the first subframe, and the input / output of the buffer 702 is the gain code of the second subframe.

The buffer 703 holds the frame erasure code B _{n + 1} of the next frame for one frame and outputs it to the prediction residual decoding unit 704, the selector 713, and the FC vector energy calculation unit 708. Since the frame erasure code output from the buffer 703 to the prediction residual decoding unit 704, the selector 713, and the FC vector energy calculation unit 708 is a frame erasure code one frame before the input frame, the frame of the current frame Erasure code _Bn .

Prediction residual decoding unit 704, logarithmic quantization prediction residual for the past M subframes (the MA prediction residual although quantized as the _{logarithm) x m-1 ~x mM,} 1 subframe previous decoding energy (Logarithmic decoding gain) e _m−1 , prediction residual bias gain e _B , next frame gain codes G _{m + 2} and G _{m + 3} , next frame frame erasure code B _{n + 1} , current frame gain codes G _m and G _{m + 1} and current frame The frame erasure code _Bn of the frame is input, and the quantization prediction residual of the current subframe is generated based on these pieces of information, and is output to the logarithmic operation unit 705 and the multiplication unit 712. Details of the prediction residual decoding unit 704 will be described later.

The logarithmic operation unit 705 calculates the logarithm of the quantized prediction residual output from the prediction residual decoding unit 704 (20 × log ₁₀ (x) in ITU-T recommendation G.729, x is an input) x _m , The data is output to the buffer 706-1.

The buffer 706-1 receives the logarithm quantized prediction residual x _m from the logarithmic operation unit 705, holds it for one subframe, and outputs it to the prediction residual decoding unit 704, the buffer 706-2, and the amplifier 707-1. . That is, the logarithmic quantization prediction residual input to these is the logarithmic quantization prediction residual x _m−1 one subframe before. Similarly, the buffers 706-i (i is 2 to M-1) hold the input logarithmic quantization prediction residual x _mi for one subframe, respectively, and the prediction residual decoding unit 704 and the buffer 706- ( i + 1) and the amplifier 707-i. The buffer 706-M holds the input log quantization prediction residual x _mM-1 for one subframe and outputs it to the prediction residual decoding unit 704 and the amplifier 707-M.

The amplifier 707-1 multiplies the logarithm quantization prediction residual x _m-1 by a predetermined MA prediction coefficient α ₁ and outputs the result to the adder 710. Similarly, the amplifier 707-j (j is 2 to M) multiplies the logarithm quantization prediction residual x _mj by a predetermined MA prediction coefficient α _j and outputs the result to the adder 710. Note that the set of MA prediction coefficients is ITU-T recommendation G.264. 729 is one type of fixed value, but a configuration in which a plurality of types of sets are prepared and an appropriate one is selected may be used.

When the frame erasure code _Bn of the current frame indicates that “the nth frame is a normal frame”, the FC vector energy calculation unit 708 calculates the energy of a separately decoded FC (fixed codebook) vector, The calculation result is output to average energy adding section 709. In addition, when the frame erasure code B _n of the current frame indicates that “the current frame is an erasure frame”, the FC vector energy calculation unit 708 sends the FC vector energy in the previous subframe to the average energy addition unit 709. Output.

The average energy addition unit 709 subtracts the FC vector energy output from the FC vector energy calculation unit 708 from the average energy, and obtains a prediction residual bias gain e _B as a subtraction result from the prediction residual decoding unit 704 and the adder. Output to 710. Here, the average energy is a constant set in advance. Energy addition and subtraction are performed in a logarithmic region.

The adder 710 calculates the sum of the logarithm quantization prediction residual after multiplication of the MA prediction coefficient output from the amplifiers 707-1 to 707-M and the prediction residual bias gain e _B output from the average energy addition unit 709. The logarithmic prediction gain as the calculation result is output to the power calculation unit 711.

The power calculation unit 711 calculates the power of the logarithmic prediction gain output from the adder 710 (10 ^x , x is an input), and outputs the prediction gain that is the calculation result to the multiplier 712.

  The multiplier 712 multiplies the prediction gain output from the power calculation unit 711 by the quantized prediction residual output from the prediction residual decoding unit 704, and outputs a decoding gain as a multiplication result to the selector 713.

The selector 713 determines the decoding gain output from the multiplier 712 based on the frame erasure code B _n of the current frame and the frame erasure code B _{n + 1} of the next frame or the decoding gain of the previous frame after attenuation output from the amplifier 715. Choose one. Specifically, when the frame erasure code _Bn of the current frame indicates that “the nth frame is a normal frame”, or when the frame erasure code _{Bn + 1 of} the next frame is “the n + 1th frame is a normal frame” If it indicates "is present", the decoding gain output from the multiplier 712 is selected, and the frame erasure code _Bn of the current frame indicates that "the nth frame is a lost frame", and When the frame erasure code B _{n + 1 of the} next frame indicates that “the (n + 1) th frame is an erasure frame”, the decoding gain of the previous frame after attenuation output from the amplifier 715 is selected. Then, the selector 713 outputs the selection result as a final decoding gain to the amplifiers 106 and 107, the buffer 714, and the logarithmic operation unit 716. When the selector 713 selects the attenuated previous frame decoding gain output from the amplifier 715, it is not actually necessary to perform all the processing from the prediction residual decoding unit 704 to the multiplier 712, and the buffer 706- Only the process of updating the contents of 1 to 706-M may be performed.

  The buffer 714 holds the decoding gain output from the selector 713 for one subframe and outputs it to the amplifier 715. As a result, the decoding gain output from the buffer 714 to the amplifier 715 is the decoding gain of one subframe before. The amplifier 715 multiplies the decoding gain of one subframe before output from the buffer 714 by a predetermined attenuation coefficient and outputs the result to the selector 713. The value of the predetermined attenuation coefficient is, for example, ITU-T recommendation G.264. However, it is sufficient that the optimum value for the codec is appropriately designed, and the value may be changed depending on the characteristics of the signal in the lost frame such as whether the lost frame is a voiced frame or an unvoiced frame.

Logarithmic operation unit 716, the selector 713 of the output decoded gain from the logarithmic (ITU-T Recommendation G.729 at _{20 × log 10 (x),} x is the input) to e _m is calculated and output to the buffer 717. Buffer 717 receives the log-decoded gain e _m from the logarithmic arithmetic unit 716, and held for one sub-frame, and outputs the prediction residual decoder 704. That is, the log decoding gain input to the prediction residual decoding unit 704 is the log decoding gain em _-1 one subframe before.

FIG. 15 is a block diagram showing an internal configuration of the prediction residual decoding unit 704 in FIG. In FIG. 15, gain codes G _m , G _{m + 1} , G _{m + 2} , and G _{m + 3} are input to the code book 801, and frame erasure codes B _n and B _{n + 1} are input to the changeover switch 812, and logarithmic quantization prediction of past M subframes residual x _m-1 ~x _mM is input to the adder 802, one sub-frame before the logarithmic decoded gain e _m-1 and the prediction residuals bias gain e _B sub-frame quantization prediction residual generation unit 807 and the sub The result is input to the frame quantization prediction residual generation unit 808.

The code book 801 decodes the corresponding quantization prediction residuals from the input gain codes G _m , G _{m + 1} , G _{m + 2} , G _{m + 3} and switches the quantization prediction residuals corresponding to the gain codes G _m , G _{m + 1.} The result is output to the changeover switch 812 via the switch 813, and the quantized prediction residual corresponding to the gain codes G _{m + 2} and G _{m + 3} is output to the logarithmic operation unit 806.

The changeover switch 813 selects one of the quantized prediction residuals decoded from the gain codes G _m and G _{m + 1} and outputs the selected quantization prediction residual to the changeover switch 812. Specifically, when performing the gain decoding processing of the first subframe selecting the quantized prediction residual decoded from gain code G _m, gain code in the case of the gain decoding processing of the second sub-frame The quantized prediction residual decoded from G _{m + 1} is selected.

The adder 802 calculates the sum of the logarithmic quantization prediction residuals x _{m−1 to} x _mM of the past M subframes, and outputs the calculation result to the amplifier 803. The amplifier 803 calculates the average value by multiplying the output value of the adder 802 by 1 / M, and outputs the calculation result to the 4 dB attenuation unit 804.

  The 4 dB attenuation unit 804 lowers the output value of the amplifier 803 by 4 dB and outputs the result to the power operation unit 805. The attenuation of 4 dB is intended to prevent the predictor from outputting an excessive prediction value in a frame (subframe) recovered from the frame loss. Therefore, in the configuration example in which such a need does not occur, the attenuator Is not necessarily required. Further, an optimum value can be freely designed for the attenuation of 4 dB.

  The power calculation unit 805 calculates the power of the output value of the 4 dB attenuation unit 804 and outputs a compensation prediction residual as a calculation result to the changeover switch 812.

The logarithmic operation unit 806 calculates the logarithm of the two quantized prediction residuals (decoded from the gain codes G _{m + 2} and G _{m + 3} ) output from the codebook 801, and calculates the logarithmic quantized prediction residual x as a calculation result. _{m + 2} and x _{m + 3} are output to subframe quantization prediction residual generation section 807 and subframe quantization prediction residual generation section 808.

The subframe quantized prediction residual generation unit 807 includes logarithmic quantized prediction residuals x _{m + 2} and x _{m + 3} , logarithmic quantized prediction residuals x _{m−1 to} x _{mM of} the past M subframes, and one subframe. enter the previous decoding energy e _m-1 and the prediction residuals bias gain e _B, calculates log quantized prediction residual of the first sub-frame based on the information, and outputs to the changeover switch 810. Similarly, the subframe quantization prediction residual generation unit 808 includes logarithmic quantization prediction residuals x _{m + 2} and x _{m + 3} , log quantization prediction residuals x _{m−1 to} x _{mM of} past M subframes, 1 subframe type a decoding energy e _m-1 and the prediction residuals bias gain e _B, calculates log quantized prediction residual of the second sub-frame based on the information, and outputs to the buffer 809. Details of the subframe quantization prediction residual generation units 807 and 808 will be described later.

The buffer 809 holds the logarithmic prediction residual of the second subframe output from the subframe quantized prediction residual generation unit 808 for one subframe, and switches the selector switch 810 when processing of the second subframe is performed. Output to. Note that, when processing the second subframe, x _{m−1 to} x _mM , e _m−1 , and e _B are updated outside the prediction residual decoding unit 704, but the subframe quantization prediction residual generation unit Neither 807 nor the subframe quantized prediction residual generation unit 808 performs any processing, and all processing is performed at the time of processing the first subframe.

  The changeover switch 810 is connected to the subframe quantization prediction residual generation unit 807 during the first subframe processing, and outputs the generated logarithmic quantization prediction residual of the first subframe to the power calculation unit 811. During the two subframe processing, the logarithm quantization prediction residual of the second subframe generated by the second subframe quantization residual generation unit 808 is connected to the buffer 809 and is output to the power calculation unit 811. The power calculation unit 811 powers the logarithmic quantization residual output from the changeover switch 810 and outputs a compensation prediction residual that is a calculation result to the changeover switch 812.

The changeover switch 812 selects the quantized prediction residual output from the codebook 801 via the changeover switch 813 when the frame erasure code _Bn of the current frame indicates “the nth frame is a normal frame”. To do. On the other hand, when the frame erasure code _Bn of the current frame indicates that “the nth frame is the erasure frame”, the changeover switch 812 indicates which information the frame erasure code _{Bn + 1} of the next frame has. To further select a compensation prediction calculation difference to be output.

That is, when the frame erasure code B _{n + 1} of the next frame indicates that “the (n + 1) th frame is a erasure frame”, the changeover switch 812 selects the compensated prediction residual output from the power calculation unit 805 and selects the next When the frame erasure code B _{n + 1} of the frame indicates that “the (n + 1) th frame is a normal frame”, the compensated prediction residual output from the power calculation unit 811 is selected. In addition, since data input to terminals other than the selected terminal is not necessary, in actual processing, it is first determined which terminal is to be selected in the changeover switch 812, and a signal output to the determined terminal is generated. It is common to perform processing for this purpose.

  FIG. 16 is a block diagram showing an internal configuration of subframe quantized prediction residual generation section 807 in FIG. Note that the internal configuration of the subframe quantization prediction residual generation unit 808 is the same as that in FIG. 16, and only the weighting coefficient value is different from that of the subframe quantization prediction residual generation unit 807.

The amplifiers 901-1 to 901 _-M multiply the input logarithmic quantization prediction residuals x _{m−1 to} x _mM by weighting coefficients β _{1 to} β _M and output the result to the adder 906. Amplifier 902, before multiplied by the weighting factor beta _-1 in log-gain e _m-1 in the subframe, and outputs to the adder 906. Amplifier 903 multiplies logarithmic bias gain e _B by weighting coefficient β _B and outputs the result to adder 906. The amplifier 904 multiplies the logarithm quantization prediction residual x _{m + 2} by the weighting coefficient β ₀₀ and outputs the result to the adder 906. The amplifier 905 multiplies the logarithm quantization prediction residual x _{m + 3} by the weighting coefficient β ₀₁ and outputs the result to the adder 906.

  Adder 906 calculates the sum of logarithmic quantization prediction residuals output from amplifiers 901-1 to 901 -M, amplifier 902, amplifier 903, amplifier 904, and amplifier 905, and outputs the calculation result to selector switch 810. .

Hereinafter, an example of how to determine the set of weighting factors β in the present embodiment will be described. As already mentioned, ITU-T Recommendation G. In the case of 729, gain quantization is subframe processing, and one frame is composed of two subframes. Therefore, the loss of one frame is a burst loss of two consecutive subframes. Therefore, the method shown in Embodiment 3 cannot determine the set of weighting factors β. Therefore, in the present embodiment, x _m and x _{m + 1} that minimize D in the following equation (6) are obtained.

Here, ITU-T Recommendation G. A case where one frame is composed of two subframes as in 729 and only one type of MA prediction coefficient is used will be described as an example. In the formula _{(6), y m-1} , y m, y m + 1, y m + 2, y m + 3, x m, x m + 1, x m + 2, x m + 3, x B, α _i is as follows.
y _m−1 : Decoding logarithmic gain y _{m + 1} of the second subframe of the previous frame y _m : Decoding logarithmic gain y _{m + 1 of the first} subframe of the current frame y Decoding logarithmic gain y _{m + 2} of the second subframe of the current frame : Decoding logarithmic gain y _{m + 3 of the} first subframe of the next frame: decoding logarithmic gain x _m of the second subframe of the next frame: logarithmic quantization prediction residual x _{m + 1} of the first subframe of the current frame: Logarithmic quantization prediction residual x _{m + 2} of the second subframe of the current frame: Logarithmic quantization prediction residual x _{m + 3 of the} first subframe of the next frame: Logarithmic quantization prediction of the second subframe of the next frame Residual x _B : Logarithmic bias gain α _i : i-th order MA prediction coefficient

And wherein the resulting equation (6) at 0 to partial differentiation for x _m, equation (6) x _{m + 1} and equation obtained at the 0 and partial differential for, as the simultaneous equations x _m and x Solving for _{m + 1} yields equations (7) and (8). Weighting coefficients β ₀₀ , β ₀₁ , β _{1 to} β _M , β ₋₁ , β _B , β ′ ₀₀ , β ′ ₀₁ , β ′ _{1 to} β ′ _M , β ′ ₋₁ , β ′ _B , are α ₀ to Since it is obtained from α _M , it is uniquely determined.

  Thus, when the next frame is normally received, the weighted addition process dedicated to the compensation process using the log quantization prediction residual received in the past and the log quantization prediction residual of the next frame is currently used. Because the logarithm quantization prediction residual of the frame is compensated, and the gain parameter is decoded using the compensated logarithm quantization prediction residual, the compensation is higher than when the past decoding gain parameter is monotonically attenuated and used. Performance can be realized.

  In addition, by using the weighting coefficient sets of Expression (7) and Expression (8) that minimize Expression (6), a lost frame (2 subframes) and a normal frame that is the next frame (2 subframes) of the lost frame It is guaranteed that the decoded logarithmic gain parameter in the frame (2 subframes) does not deviate significantly from the logarithmic gain parameter of the previous subframe of the lost frame. For this reason, even if the decoding logarithmic gain parameter of the next frame (2 subframes) is unknown, the reception information (logarithm quantization prediction residual) of the next frame (2 subframes) is used effectively, The risk (compensation that deviates greatly from the correct decoding gain parameter) in the case of compensation in a different direction can be minimized.

(Embodiment 5)
FIG. 17 is a block diagram showing the main configuration of a speech encoding apparatus according to Embodiment 5 of the present invention. FIG. 17 illustrates an example in which the weighting coefficient set is determined by the second method described in the third embodiment and the compensation mode information En _{+ 1} is encoded, that is, the MA prediction coefficient mode information of the nth frame is used. A method of expressing compensation mode information of n−1 frames with 1 bit will be described.

In this case, the previous frame LPC compensation unit 1003 uses the weighted sum of the decoded quantized prediction residual of the current frame and the decoded quantized predicted residual of M + 1 frames before 2 frames as described with reference to FIG. The compensation LSF of the (n-1) th frame is obtained. In FIG. 13, the compensation LSF of the nth frame is obtained using the encoding information of the (n + 1) th frame, but here, the compensation LSF of the (n−1) th frame is obtained using the encoding information of the nth frame. Therefore, the frame number is shifted by one. In other words, the combination of α _i ^(j) and α ′ _i ^(j) is limited to two out of four combinations by the MA prediction coefficient code of the nth frame (= current frame) (that is, the MA of the nth frame). When the prediction mode is mode 0, the combination of the MA prediction modes of the (n-1) th frame and the (n-1) th frame is either (0-0) or (1-0). The previous frame LPC compensation unit 1003 generates two types of compensation LSF ω0 _n ^(j) and ω1 _n ^(j) using the set of two types of weighting coefficients β.

Compensation mode determiner 1004, the determination of the mode based on whether near .omega.0 _n ^(j) and ω1 which _n of the ^(j) is the input LSF omega _n ^(j). The degree of separation between ω 0 _n ^(j) and ω 1 _n ^(j) and ω _n ^(j) may be based on a simple Euclidean distance or ITU-T Recommendation G. It may be based on a weighted Euclidean distance as used in L729 quantization of 729.

  Hereinafter, the operation of each unit of the speech encoding apparatus in FIG. 17 will be described.

Input signal s _n is, LPC analyzing section 1001, it is input to the target vector calculation unit 1006 and a filter state update unit 1013.

LPC analyzing section 1001 performs a known linear prediction analysis on the input signal s _n, the linear prediction coefficients a _{j (j} = 0~M, M is a linear prediction analysis order .a ₀ = 1.0) The impulse response calculator 1005, output to the target vector calculator 1006 and LPC encoder 1002. Further, the LPC analysis unit 1001 converts the linear prediction coefficient a _j into the LSF parameter ω _n ^(j) and outputs it to the compensation mode determination unit 1004.

The LPC encoding unit 1002 quantizes and encodes the input LPC (linear prediction coefficient) and sends the quantized linear prediction coefficient a ′ _j to the impulse response calculation unit 1005, the target vector calculation unit 1006, and the synthesis filter unit 1011. Output. In this example, LPC quantization and coding are performed in the LSF parameter area. Further, the LPC encoding unit 1002 outputs the LPC encoding result L _n to the multiplexing unit 1014, and the quantization prediction residual x _n , the decoded quantization LSF parameter ω ′ _n ^(j), and the MA prediction quantization mode. and outputs the K _n before frame LPC compensator 1003.

The previous frame LPC compensation unit 1003 holds the decoded quantized LSF parameter ω ′ _n ^(j) of the nth frame output from the LPC encoding unit 1002 in the buffer for two frames. The decoded quantized LSF parameter two frames before is ω ′ _n−2 ^(j) . Also, the previous frame LPC compensation unit 1003 holds the decoded quantized prediction residual x _n of the nth frame for M + 1 frames. The previous frame LPC compensation unit 1003 also performs the quantization prediction residual x _n , the decoded quantization LSF parameter ω ′ _n−2 ^(j) two frames before, and the decoded quantization prediction residual M + 1 frames before the second frame. The decoded quantized LSF parameters ω0 _n ^(j) and ω1 _n ^(j) of the ⁽ n−1 ⁾ _th frame are generated by the weighted sum of x _{n−2 to} x _nM−1 and output to the compensation mode determiner 1004. Here, the previous frame LPC compensator 1003 is provided with four sets of weighting coefficients for obtaining the weighted sum, MA predictive quantization mode information K _n input from LPC encoding section 1002 or 0 Depending on whether it is 1, two of the four types are selected and used to generate ω0 _n ^(j) and ω1 _n ^(j) .

The compensation mode determiner 1004 outputs either of the two types of compensation LSF parameters ω 0 _n ^(j) and ω 1 _n ^(j) output from the previous frame LPC compensation unit 1003 and the unquantized LSF parameter ω output from the LPC analysis unit 1001. determine close to _n ^(j), and outputs a code E _n to multiplexing section 1014 corresponding to the set of weighting coefficients for generating a compensation LSF parameters closer.

The impulse response calculation unit 1005 uses the unquantized linear prediction coefficient a _j output from the LPC analysis unit 1001 and the quantized linear prediction coefficient a ′ _j output from the LPC encoding unit 1002 to generate impulses of the perceptual weighting synthesis filter. A response h is generated and output to the ACV encoding unit 1007 and the FCV encoding unit 1008.

The target vector calculation unit 1006 includes the input signal s _n , the unquantized linear prediction coefficient a _j output from the LPC analysis unit 1001, the quantized linear prediction coefficient a ′ _j output from the LPC encoding unit 1002, and the filter state update. A target vector (a signal obtained by removing the zero input response of the perceptual weighting synthesis filter from the signal obtained by applying the perceptual weighting filter to the input signal) o from the filter state output from the units 1012, 1013, and an ACV encoding unit 1007; The result is output to gain encoding section 1009 and filter state updating section 1012.

The ACV encoding unit 1007 receives the target vector o from the target vector calculation unit 1006, the impulse response h of the auditory weighting synthesis filter from the impulse response calculation unit 1005, and the excitation signal ex generated in the previous frame from the excitation generation unit 1010. and performs adaptive codebook search, adaptive codebook code _{a n} is the result to the multiplexing unit 1014, a quantization pitch lag T to FCV encoding section 1008, the AC vector v to the sound source generating unit 1010, AC vector v The filtered AC vector component p obtained by convolving the impulse response h of the auditory weighting synthesis filter with the filter state update unit 1012 and the gain encoding unit 1009 is FCV-encoded for the target vector o ′ updated for the fixed codebook search. Output to the unit 1008. A more specific search method is ITU-T recommendation G.264. 729 and the like. Although omitted in FIG. 17, it is general to reduce the amount of calculation required for the adaptive codebook search by determining the range for performing the closed loop pitch search by open loop pitch search or the like.

The FCV encoding unit 1008 receives the fixed codebook target vector o ′ and the quantization pitch lag T from the ACV encoding unit 1007, and the impulse response h of the auditory weighting synthesis filter from the impulse response calculation unit 1005, respectively. -T Recommendation G. By methods such as those described in 729 performs a fixed codebook search, the fixed codebook code F _n to the multiplexing unit 1014, the FC vector u to the sound source generating unit 1010, the impulse response of the perceptually weighted filter to the FC vector u Are output to the filter state update unit 1012 and the gain encoding unit 1009, respectively.

Gain encoding section 1009 receives target vector o from target vector calculation section 1006, AC vector component p after filtering from ACV encoding section 1007, and FC vector component q after filtering from FCV encoding section 1008. Then, a set of ga and gf having the minimum | o− (ga × p + gf × q) | ² is output to the sound source generation unit 1010 as a quantization adaptive codebook gain and a quantization fixed codebook gain.

  The excitation generator 1010 generates an adaptive codebook vector v from the ACV encoding unit 1007, a fixed codebook vector u from the FCV encoding unit 1008, and an adaptive codebook vector gain ga and a fixed codebook vector gain from the gain encoding unit 1009. gf is input, the excitation vector ex is calculated by ga × v + gf × u, and is output to the ACV encoding unit 1007 and the synthesis filter unit 1011. The excitation vector ex output to the ACV encoding unit 1007 is used to update the ACB (excitation vector buffer generated in the past) in the ACV encoding unit.

The synthesis filter unit 1011 drives a linear prediction filter composed of the quantized linear prediction coefficient a ′ _j output from the LPC encoding unit 1002 using the excitation vector ex output from the excitation generation unit 1010, and locally decoded speech A signal s ′ _n is generated and output to the filter state update unit 1013.

  The filter state update unit 1012 receives the composite adaptive codebook vector p from the ACV encoding unit 1007, the composite fixed codebook vector q from the FCV encoding unit 1008, and the target vector o from the target vector calculation unit 1006, respectively. The filter state of the auditory weighting filter in the target vector calculation unit 1006 is generated and output to the target vector calculation unit 1006.

Filter state update unit 1013, the error between the local decoded speech s' _n output from the synthesizing filter unit 1011 and the input signal s _n is calculated, the target vector calculated as a state of the synthesis filter in the target vector calculation section 1006 Output to the unit 1006.

Multiplexing unit 1014, a code _{F n, A n, G n} , L n, and outputs the multiplexed coded information _{E n.}

  Further, in the present embodiment, an example in which an error from the unquantized LSF parameter is calculated only for the decoded quantized LSF parameter of the (n−1) th frame is shown, but the decoded quantized LSF parameter of the nth frame and the nth frame The compensation mode may be determined in consideration of an error from the unquantized LSF parameter.

  Thus, according to the speech coding apparatus according to the present embodiment, a weighting coefficient set for compensation processing that is optimal for compensation processing is specified in correspondence with the speech decoding device of Embodiment 3, and the information Is transmitted to the decoder side, higher compensation performance is obtained on the decoder side, and the quality of the decoded speech signal is improved.

(Embodiment 6)
FIG. 18 is a block diagram showing configurations of an audio signal transmitting apparatus and an audio signal receiving apparatus that constitute an audio signal transmission system according to Embodiment 6 of the present invention. The only difference is that the speech coding apparatus according to the fifth embodiment is applied to the speech signal transmitting apparatus and the speech decoding apparatus according to any one of the first to third embodiments is applied to the speech signal receiving apparatus.

  The audio signal transmission apparatus 1100 includes an input apparatus 1101, an A / D conversion apparatus 1102, an audio encoding apparatus 1103, a signal processing apparatus 1104, an RF modulation apparatus 1105, a transmission apparatus 1106, and an antenna 1107.

  An input terminal of the A / D conversion device 1102 is connected to the input device 1101. The input terminal of the speech encoding device 1103 is connected to the output terminal of the A / D conversion device 1102. The input terminal of the signal processing device 1104 is connected to the output terminal of the speech encoding device 1103. The input terminal of the RF modulation device 1105 is connected to the output terminal of the signal processing device 1104. An input terminal of the transmission device 1106 is connected to an output terminal of the RF modulation device 1105. The antenna 1107 is connected to the output terminal of the transmission device 1106.

  The input device 1101 receives an audio signal, converts it into an analog audio signal, which is an electrical signal, and provides the analog audio signal to the A / D conversion device 1102. The A / D converter 1102 converts an analog voice signal from the input device 1101 into a digital voice signal, and provides this to the voice encoder 1103. The audio encoding device 1103 encodes the digital audio signal from the A / D conversion device 1102 to generate an audio encoded bit string, and provides the generated signal to the signal processing device 1104. The signal processing device 1104 performs channel coding processing, packetization processing, transmission buffer processing, and the like on the speech coded bit sequence from the speech coding device 1103, and then provides the speech coded bit sequence to the RF modulation device 1105. The RF modulation device 1105 modulates the speech encoded bit string signal subjected to the channel coding processing and the like from the signal processing device 1104 and supplies the modulated signal to the transmission device 1106. The transmission device 1106 transmits the modulated audio encoded signal from the RF modulation device 1105 as a radio wave (RF signal) via the antenna 1107.

  In the audio signal transmitting apparatus 1100, the digital audio signal obtained via the A / D conversion apparatus 1102 is processed in units of several tens of frames. When the network constituting the system is a packet network, encoded data of one frame or several frames is put into one packet and the packet is transmitted to the packet network. When the network is a circuit switching network, packetization processing and transmission buffer processing are not necessary.

  The audio signal reception device 1150 includes an antenna 1151, a reception device 1152, an RF demodulation device 1153, a signal processing device 1154, an audio decoding device 1155, a D / A conversion device 1156, and an output device 1157.

  An input terminal of the reception device 1152 is connected to the antenna 1151. The input terminal of the RF demodulator 1153 is connected to the output terminal of the receiver 1152. Two input terminals of the signal processor 1154 are connected to two output terminals of the RF demodulator 1153. Two input terminals of the speech decoding apparatus 1155 are connected to two output terminals of the signal processing apparatus 1154. The input terminal of the D / A conversion device 1156 is connected to the output terminal of the speech decoding device 1155. The input terminal of the output device 1157 is connected to the output terminal of the D / A converter 1156.

  Receiving device 1152 receives a radio wave (RF signal) including speech coding information via antenna 1151, generates a received speech coded signal that is an analog electrical signal, and provides this to RF demodulating device 1153. The radio wave (RF signal) received via the antenna is exactly the same as the radio wave (RF signal) sent out by the audio signal transmitting device unless there is signal attenuation or noise superposition in the transmission path.

  The RF demodulator 1153 demodulates the received speech encoded signal from the receiver 1152 and provides it to the signal processor 1154. Further, information indicating whether or not the received speech encoded signal has been demodulated normally is given to the signal processing device 1154 separately. The signal processing device 1154 performs jitter absorption buffering processing, packet assembling processing, channel decoding processing, and the like of the received speech encoded signal from the RF demodulating device 1153, and provides the received speech encoded bit string to the speech decoding device 1155. . Also, information indicating whether or not the received speech encoded signal was successfully demodulated is input from the RF demodulator 1153, and the information input from the RF demodulator 1153 indicates that “the demodulator could not be demodulated normally”, or When the received speech encoded bit string cannot be normally decoded without normally performing packet assembly processing or the like in the signal processing device, the fact that frame loss has occurred is given to the speech decoding device 1155 as frame loss information. The audio decoding device 1155 performs a decoding process on the received audio encoded bit string from the signal processing device 1154 to generate a decoded audio signal and supplies it to the D / A conversion device 1156. Speech decoding apparatus 1155 determines whether to perform normal decoding processing or to perform decoding processing by frame loss compensation (concealment) processing in accordance with frame loss information input in parallel with the received speech encoded bit string. The D / A conversion device 1156 converts the digital decoded speech signal from the speech decoding device 1155 into an analog decoded speech signal and gives it to the output device 1157. The output device 1157 converts the analog decoded audio signal from the D / A converter 1156 into air vibration and outputs it as sound waves so that it can be heard by the human ear.

  As described above, even when a transmission path error (particularly, a frame loss error typified by packet loss) occurs by providing the speech encoding device and speech decoding device described in Embodiments 1 to 5, It is possible to obtain a decoded speech signal with better quality than before.

(Embodiment 7)
In the first to sixth embodiments, the case where the MA type is used as the prediction model has been described. However, the present invention is not limited to this, and the AR type can also be used as the prediction model. In Embodiment 7, a case where an AR type is used as a prediction model will be described. The configuration of the speech decoding apparatus according to Embodiment 7 is the same as that shown in FIG. 1 except that the internal configuration of the LPC decoding unit is different.

  FIG. 19 is a block diagram showing an internal configuration of LPC decoding section 105 of the speech decoding apparatus according to the present embodiment. In FIG. 19, the same components as those in FIG. 2 are denoted by the same reference numerals as those in FIG. 2, and detailed description thereof is omitted.

  Compared with FIG. 2, the LPC decoding unit 105 shown in FIG. 19 includes a part related to prediction (buffer 204, amplifier 205, adder 206) and a part related to frame erasure compensation (code vector decoding unit 203, buffer 207). A configuration is adopted in which components that are deleted and replaced (code vector decoding unit 1901, amplifier 1902, adder 1903, and buffer 1904) are added.

The LPC code L _{n + 1} is input to the buffer 201 and the code vector decoding unit 1901, and the frame erasure code B _{n + 1} is input to the buffer 202, the code vector decoding unit 1901 and the selector 209.

The buffer 201 holds the LPC code L _{n + 1} of the next frame for one frame and outputs it to the code vector decoding unit 1901. The LPC code output from the buffer 201 to the code vector decoding unit 1901 becomes the LPC code L _n of the current frame as a result of being held in the buffer 201 for one frame.

The buffer 202 holds the frame erasure code B _{n + 1} of the next frame for one frame and outputs it to the code vector decoding unit 1901. The frame erasure code output from the buffer 202 to the code vector decoding unit 1901 becomes the frame erasure code _Bn of the current frame as a result of being held in the buffer 202 for one frame.

The code vector decoding unit 1901 includes a decoded LSF vector y _{n-1 of} the previous frame, an LPC code L _{n + 1 of} the next frame, a frame erasure code B _{n + 1 of} the next frame, an LPC code L _n of the current frame, and a frame erasure code of the current frame B _n is input, a quantized prediction residual vector x _n of the current frame is generated based on these pieces of information, and output to the adder 1903. Details of the code vector decoding unit 1901 will be described later.

The amplifier 1902 multiplies the decoded LSF vector y _n−1 of the previous frame by a predetermined AR prediction coefficient a ₁ and outputs the result to the adder 1903.

The adder 1903 outputs the prediction LSF vector output from the amplifier 1902 (that is, the decoded LSF vector of the previous frame multiplied by the AR prediction coefficient) and the quantized prediction residual vector of the current frame output from the code vector decoding unit 1901. The sum with x _n is calculated, and the decoded LSF vector y _n as the calculation result is output to the buffer 1904 and the LPC converter 208.

The buffer 1904 holds the decoded LSF vector y _n of the current frame for one frame and outputs it to the code vector decoding unit 1901 and the amplifier 1902. The decoded LSF vector input to these becomes the decoded LSF vector y _{n-1 of the} previous frame as a result of being held for one frame in the buffer 1904.

  Note that when the selector 209 selects the decoded LPC parameter in the previous frame output from the buffer 210, it is not necessary to actually perform all the processes from the code vector decoding unit 1901 to the LPC conversion unit 208.

  Next, the internal configuration of the code vector decoding unit 1901 in FIG. 19 will be described in detail with reference to the block diagram in FIG.

The code book 2001 generates a code vector specified by the LPC code L _n of the current frame and outputs the code vector to the changeover switch 309, and generates a code vector specified by the LPC code L _{n + 1 of} the next frame and supplies it to the amplifier 2002. Output. The code book may have a multi-stage configuration or a split configuration.

The amplifier 2002 multiplies the code vector x _{n + 1} output from the code book 2001 by the weighting coefficient b ₀ and outputs the result to the adder 2005.

The amplifier 2003 performs a process of obtaining a quantized prediction residual vector in the current frame necessary for generating the decoded LSF vector of the previous frame. That is, the amplifier 2003 calculates the vector x _{n of the} current frame so that the decoded LSF vector y _n−1 of the previous frame becomes the decoded LSF vector y _n of the current frame. Specifically, the amplifier 2003 multiplies the input decoded LSF vector y _n-1 of the previous frame by a coefficient (1-a ₁ ). Then, the amplifier 2003 outputs the calculation result to the changeover switch 309.

The amplifier 2004 multiplies the input decoded LSF vector y _n−1 of the previous frame by the weighting coefficient b ₋₁ and outputs the result to the adder 2005.

The adder 2005 calculates the sum of the vectors output from the amplifier 2002 and the amplifier 2004, and outputs the code vector that is the calculation result to the changeover switch 309. That is, adder 2005 calculates vector _xn of the current frame by weighted addition of the code vector specified by LPC code Ln _{+ 1} of the next frame and the decoded LSF vector of the previous frame.

When the frame erasure code _Bn of the current frame indicates that “the nth frame is a normal frame”, the changeover switch 309 selects the code vector output from the codebook 2001 and quantizes the code frame of the current frame. The prediction residual vector x _n is output. On the other hand, when the switch 309 indicates that the frame erasure code B _n of the current frame indicates “the nth frame is a erasure frame”, which information the frame erasure code B _{n + 1} of the next frame has. To further select a vector to be output.

That is, when the frame erasure code B _{n + 1 of} the next frame indicates that “the (n + 1) th frame is an erasure frame”, the changeover switch 309 selects the vector output from the amplifier 2003 and uses this as the quantum of the current frame. Output as a generalized prediction residual vector x _n . In this case, it is not necessary to perform the process of generating vectors from the code book 2001 and the amplifiers 2002 and 2004 to the adder 2005. In this case, since y _n−1 may be used as y _n , x _n is not necessarily generated by the processing of the amplifier 2003.

When the frame erasure code B _{n + 1 of} the next frame indicates that “the (n + 1) th frame is a normal frame”, the changeover switch 309 selects the vector output from the adder 2005 and replaces it with the current frame. The quantized prediction residual vector x _n is output. In this case, it is not necessary to perform the processing of the amplifier 2003.

Note that the compensation processing of the present embodiment is such that the distance between the decoding parameter y _n-1 of the n−1th frame and the decoding parameter y _n of the nth frame, , The weighting coefficient b ₋₁ and the weighting coefficient b ₋₁ so that the sum D of the distances between the decoding parameter y _n of the nth frame and the decoding parameter y _{n + 1} of the ( _{n + 1) th} frame becomes small. b ₀ is determined.

Hereinafter, an example of how to determine the weighting coefficients b ₋₁ and b ₀ will be described. In order to minimize D in equation (9), the following equation (10) is solved for the decoded quantized prediction residual x _n of the lost nth frame. As a result, x _n can be obtained by the following equation (11). Note that when the prediction coefficient differs in each order, Equation (9) is replaced with Equation (12). a ₁ is the AR prediction coefficient, and a ₁ ^(j) is multiplied by the ^j _-th component of the AR prediction coefficient set (ie, y _n-1 ^(j) which is the j-th component of the decoded LSF vector y _n-1 of the previous frame ⁾ . Coefficient).

X, y and a in the above formula are as follows.
x _n ^(j) : the quantized prediction residual y _n ^(j) of the LSF parameter in the nth frame: ^j _n component a ₁ ^(j) of the decoded LSF parameter in the nth frame: of the AR prediction coefficient set Jth component

  Thus, according to the present embodiment using the AR type as a prediction model, if the next frame is normally received when the current frame is lost, the parameters decoded in the past and the next frame Compensation processing of the decoded quantized prediction residual of the LSF parameter of the current frame is performed by weighted addition processing (weighted linear sum) dedicated to the compensation processing using the quantized prediction residual, and LSF is used using the compensated quantized prediction residual. Decode parameters. As a result, it is possible to realize higher compensation performance than repeatedly using past decoded LSF parameters.

  Note that the contents described in Embodiments 2 to 4 can be applied to this embodiment using the AR type, and even in this case, the same effect as described above can be obtained.

(Embodiment 8)
In the seventh embodiment, the case where there is only one set of prediction coefficients has been described. However, the present invention is not limited to this, and the case where there are a plurality of sets of prediction coefficients as in the second and third embodiments. Can be applied. In the eighth embodiment, an example in which an AR type prediction model having a plurality of types of prediction coefficient sets is used will be described.

FIG. 21 is a block diagram of the speech decoding apparatus according to the eighth embodiment. The configuration of speech decoding apparatus 100 shown in FIG. 21 is different except that the internal configuration of the LPC decoding unit is different and that there is no input line for compensation mode information En _{+ 1} from demultiplexing unit 101 to LPC decoding unit 105. , The same as FIG.

  FIG. 22 is a block diagram showing an internal configuration of LPC decoding section 105 of the speech decoding apparatus according to this embodiment. In FIG. 22, the same components as those in FIG. 19 are denoted by the same reference numerals as those in FIG. 19, and detailed description thereof will be omitted.

  The LPC decoding unit 105 illustrated in FIG. 22 employs a configuration in which a buffer 2202 and a coefficient decoding unit 2203 are added as compared with FIG. Further, the operation and internal configuration of the code vector decoding unit 2201 in FIG. 22 are different from those in the code vector decoding unit 1901 in FIG.

The LPC code V _{n + 1} is input to the buffer 201 and the code vector decoding unit 2201, and the frame erasure code B _{n + 1} is input to the buffer 202, the code vector decoding unit 2201 and the selector 209.

The buffer 201 holds the LPC code V _{n + 1} of the next frame for one frame and outputs it to the code vector decoding unit 2201. The LPC code output from the buffer 201 to the code vector decoding unit 2201 becomes the LPC code V _n of the current frame as a result of being held in the buffer 201 for one frame. The buffer 202 holds the frame erasure code B _{n + 1} of the next frame for one frame and outputs it to the code vector decoding unit 2201.

The code vector decoding unit 2201 includes a decoded LSF vector y _{n−1 of} the previous frame, an LPC code V _{n + 1 of} the next frame, a frame erasure code B _{n + 1 of} the next frame, an LPC code V _n of the current frame, and a prediction coefficient code of the next frame K _{n + 1} and the frame erasure code B _n of the current frame are input, a quantized prediction residual vector x _n of the current frame is generated based on these information, and output to the adder 1903. Details of the code vector decoding unit 2201 will be described later.

The buffer 2202 holds the AR prediction coefficient code K _{n + 1} for one frame and outputs it to the coefficient decoding unit 2203. Consequently, AR prediction coefficient code output from the buffer 2202 to the coefficient decoding section 2203, a previous frame of the AR prediction coefficient code _{K n.}

Coefficient decoding section 2203 stores a plurality of types of coefficient sets, and identifies coefficient sets by frame erasure codes B _n and B _{n + 1} and AR prediction coefficient codes K _n and K _{n + 1} . Here, the coefficient decoding unit 2203 specifies the coefficient set in the following three ways.

If the frame erasure code B _n input indicates that "the n-th frame is normal frame", coefficient decoding portion 2203 selects the coefficient set specified by AR prediction coefficient code K _n.

When the input frame erasure code B _n indicates that “the nth frame is a erasure frame” and the frame erasure code B _{n + 1} indicates that “the (n + 1) th frame is a normal frame”, the coefficient decoding unit 2203 The coefficient set to be selected is determined using the AR prediction coefficient code K _{n + 1} received as the parameter of the (n + 1) th frame. That is, K _{n + 1} is used as it is instead of the AR prediction coefficient code K _n . Alternatively, a coefficient set to be used in such a case may be determined in advance, and the determined coefficient set may be used regardless of K _{n + 1} .

Further, when the input frame erasure code _Bn indicates that “the nth frame is a erasure frame” and the frame erasure code _{Bn + 1} indicates that “the n + 1th frame is an erasure frame”, information that can be used Is only information on the coefficient set used in the previous frame, the coefficient decoding unit 2203 repeatedly uses the coefficient set used in the previous frame. Alternatively, a predetermined coefficient set may be used in a fixed manner.

Then, the coefficient decoding unit 2203 outputs the AR prediction coefficient a ₁ to the amplifier 1902 and outputs the AR prediction coefficient (1-a ₁ ) to the code vector decoding unit 2201.

The amplifier 1902 multiplies the decoded LSF vector y _n−1 of the previous frame by the AR prediction coefficient a ₁ input from the coefficient decoding unit 2203 and outputs the result to the adder 1903.

  Next, the internal configuration of the code vector decoding unit 2201 in FIG. 22 will be described in detail with reference to the block diagram in FIG. In FIG. 23, components common to those in FIG. 20 are denoted by the same reference numerals as those in FIG. 20, and detailed descriptions thereof are omitted. 23 employs a configuration in which a coefficient decoding unit 2301 is added to the code vector decoding unit 1901 in FIG.

The coefficient decoding unit 2301 stores a plurality of types of coefficient sets, specifies the coefficient set by the AR prediction coefficient code K _{n + 1} , and outputs the coefficient set to the amplifiers 2002 and 2004. The coefficient set used here can be calculated using the AR prediction coefficient a ₁ output from the coefficient decoding unit 2203. In this case, it is not necessary to store the coefficient set, and the AR prediction is performed. The coefficient a ₁ may be input and calculated. A specific calculation method will be described later.

The code book 2001 generates a code vector specified by the LPC code V _n of the current frame and outputs the code vector to the changeover switch 309, and generates a code vector specified by the LPC code V _{n + 1 of} the next frame and supplies it to the amplifier 2002. Output. The code book may have a multi-stage configuration or a split configuration.

The amplifier 2002 multiplies the code vector x _{n + 1} output from the code book 2001 by the weighting coefficient b ₀ output from the coefficient decoding unit 2301, and outputs the result to the adder 2005.

The amplifier 2003 multiplies the decoded LSF vector y _n-1 of the previous frame by the AR prediction coefficient (1-a ₁ ) output from the coefficient decoding unit 2203, and outputs the result to the changeover switch 309. It is to be noted that the output of the buffer 1904 can be input to the LPC conversion unit 208 in place of the output of the adder 1903 instead of performing the processing of the amplifier 2003, the amplifier 1902, and the adder 1903 without creating such a path in implementation. If a simple switching configuration is provided, a route through the amplifier 2003 is not necessary.

The amplifier 2004 multiplies the input decoded LSF vector y _n−1 of the previous frame by the weighting coefficient b ₋₁ output from the coefficient decoding unit 2301, and outputs the result to the adder 2005.

Note that the compensation processing of the present embodiment is such that the distance between the decoding parameter y _n-1 of the n−1th frame and the decoding parameter y _n of the nth frame, , The weighting coefficient b ₋₁ and the weighting coefficient b ₋₁ so that the sum D of the distances between the decoding parameter y _n of the nth frame and the decoding parameter y _{n + 1} of the ( _{n + 1) th} frame becomes small. b ₀ is determined.

Hereinafter, an example of how to determine the weighting coefficients b ₋₁ and b ₀ will be described. In order to minimize D in equation (13), the following equation (14) is solved for the decoded quantized prediction residual x _n of the lost nth frame. As a result, x _n can be obtained by the following equation (15). In addition, when a prediction coefficient differs in each order, Formula (13) replaces Formula (16). a ′ ₁ is the AR prediction coefficient in the (n + 1 ⁾ th frame, a ₁ is the AR prediction coefficient in the nth frame, a ₁ ^(j) is the jth component of the AR prediction coefficient set (ie, the decoded LSF vector y _{n−1 of the} previous frame) Represents a coefficient to be multiplied by y _n-1 ^(j) which is the j-th component of.

X, y and a in the above formula are as follows.
x _n ^(j) : quantized prediction residual y _n ^(j) of the LSF parameter in the nth frame: ^j j component a ₁ ^(j) of the decoded LSF parameter in the nth frame: AR of the nth frame J-th component a ′ ₁ ^{(j) of} prediction coefficient set: j-th component of AR prediction coefficient set of the ⁽ n + 1 ⁾ th frame

Here, if the nth frame is an erasure frame, the prediction coefficient set of the nth frame is unknown. how to determine the a ₁ is considered some. First, there is a method of sending as additional information in the (n + 1) th frame as in the second embodiment. However, an additional bit is required and correction is also required on the encoder side. Next, there is a method of using the prediction coefficient set used in the (n-1) th frame. Furthermore, there is a method of using the prediction coefficient set received in the (n + 1) th frame. In this case, a ₁ = a ′ ₁ . Furthermore, there is a method that always uses a specific set of prediction coefficients. However, as will be described later, even if different a ₁ is used, if AR prediction is performed using the same a ₁ , the decoded y _n becomes equal. In the case of predictive quantization using AR prediction, the quantized prediction residual x _n is not related to prediction, and only the decoded quantization parameter y _n is related to prediction. In this case, a ₁ may be an arbitrary value. .

If a ₁ is determined, b ₀ and b ₁ can be determined from Expression (15) or Expression (16), and the code vector x _n of the lost frame can be generated.

Incidentally, substituting code vector x _n of the lost frame obtained by the above equation (16) into equation _{(y n = a 1 y n} -1 + x n) representing the y _n, as in the following formula (17) Become. Therefore, the decoding parameters in the lost frame generated by the compensation process can be directly obtained from x _{n + 1} , y _n−1, and a ₁ ′. In this case, compensation processing that does not use the prediction coefficient a ₁ in the lost frame is possible.

  Thus, according to the present embodiment, in addition to the features described in the seventh embodiment, a plurality of sets of prediction coefficients are prepared and the compensation process is performed, so that a higher compensation performance than in the seventh embodiment is achieved. Can be realized.

(Embodiment 9)
In the first to eighth embodiments, the case where n frames are decoded after receiving n + 1 frames has been described. However, the present invention is not limited to this, and the generation of n frames is performed using decoding parameters of n−1 frames. It is also possible to perform n-frame parameter decoding using the method of the present invention at the time of decoding n + 1 frame, and update the internal state of the predictor with the result, thereby decoding n + 1 frame.

  Embodiment 9 describes this case. The configuration of the speech decoding apparatus according to the ninth embodiment is the same as that shown in FIG. The configuration of the LPC decoding unit 105 may be the same as that shown in FIG. 19, but it is rewritten as shown in FIG. 24 in order to clarify that n + 1 frame decoding is performed for n + 1 frame encoded information input.

  FIG. 24 is a block diagram showing an internal configuration of LPC decoding section 105 of the speech decoding apparatus according to the present embodiment. 24, the same reference numerals as those in FIG. 19 are given to the same components as those in FIG. 19, and detailed description thereof will be omitted.

Compared with FIG. 19, the LPC decoding unit 105 shown in FIG. 24 deletes the buffer 201, the output of the code vector decoding unit is x _{n + 1} , and the decoding parameter is n + 1 frame (y _{n + 1} ). And a configuration in which a changeover switch 2402 is added. 24 is different from the code vector decoding unit 1901 in FIG. 19 in the operation and internal configuration of the code vector decoding unit 2401 in FIG.

The LPC code L _{n + 1} is input to the code vector decoding unit 2401, and the frame erasure code B _{n + 1} is input to the buffer 202, the code vector decoding unit 2401, and the selector 209.

The buffer 202 holds the frame erasure code B _{n + 1} of the current frame for one frame and outputs it to the code vector decoding unit 2401. The frame erasure code output from the buffer 202 to the code vector decoding unit 2401 becomes the frame erasure code _Bn of the previous frame as a result of being held in the buffer 202 for one frame.

The code vector decoding unit 2401 receives the decoded LSF vector y _{n−1 of} the previous frame, the LPC code L _{n + 1 of} the current frame, and the frame erasure code B _{n + 1} of the current frame, and quantizes the current frame based on these information The prediction residual vector x _{n + 1} and the decoded LSF vector y ′ _n of the previous frame are generated and output to the adder 1903 and the changeover switch 2402, respectively. Details of the code vector decoding unit 2401 will be described later.

The amplifier 1902 multiplies the decoded LSF vector y _n or y ′ _n of the previous frame by a predetermined AR prediction coefficient a ₁ and outputs the result to the adder 1903.

The adder 1903 calculates the predicted LSF vector output from the amplifier 1902 (that is, the product of the decoded LSF vector of the previous frame multiplied by the AR prediction coefficient), and the decoded LSF vector y _{n + 1} that is the calculation result is stored in the buffer 1904 and The data is output to the LPC conversion unit 208.

The buffer 1904 holds the decoded LSF vector y _{n + 1} of the current frame for one frame and outputs it to the code vector decoding unit 2401 and the changeover switch 2402. The decoded LSF vector input to these becomes the decoded LSF vector y _n one frame before as a result of being held for one frame in the buffer 1904.

The changeover switch 2402 is either the decoded LSF vector y _{n of} the previous frame or the decoded LSF vector y ′ _{n of} the previous frame regenerated using the LPC code L _{n + 1} of the current frame by the code vector decoding unit 2401. Is selected by the frame erasure code B _n of the previous frame. The changeover switch 2402 selects y ′ _n when B _n indicates a lost frame.

  Note that when the selector 209 selects the decoded LPC parameter in the previous frame output from the buffer 210, it is not necessary to actually perform all the processes from the code vector decoding unit 2401 to the LPC conversion unit 208.

  Next, the internal configuration of the code vector decoding unit 2401 in FIG. 24 will be described in detail with reference to the block diagram in FIG. In FIG. 25, components common to those in FIG. 20 are denoted by the same reference numerals as those in FIG. 20, and detailed descriptions thereof are omitted. The code vector decoding unit 2401 in FIG. 25 employs a configuration in which a buffer 2502, an amplifier 2503, and an adder 2504 are added to the code vector decoding unit 1901 in FIG. Also, the operation and internal configuration of the selector switch 2501 in FIG. 25 are different from those in the selector switch 309 in FIG.

The code book 2001 generates a code vector specified by the LPC code L _{n + 1} of the current frame, outputs the code vector to the changeover switch 2501, and outputs the code vector to the amplifier 2002.

The amplifier 2003 performs a process of obtaining a quantized prediction residual vector in the current frame necessary for generating the decoded LSF vector of the previous frame. That is, the amplifier 2003 calculates the vector x _{n + 1} of the current frame so that the decoded LSF vector y _n of the previous frame becomes the decoded LSF vector y _{n + 1} of the current frame. Specifically, the amplifier 2003 multiplies the input decoded LSF vector y _n of the previous frame by a coefficient (1-a ₁ ). Then, the amplifier 2003 outputs the calculation result to the changeover switch 2501.

When the frame erasure code B _{n + 1} of the current frame indicates that “the (n + 1) th frame is a normal frame”, the changeover switch 2501 selects the code vector output from the code book 2001, and selects this code vector. The quantized prediction residual vector x _{n + 1} is output.
On the other hand, if the frame erasure code B _{n + 1} of the current frame indicates that “the (n + 1) th frame is a erasure frame”, the changeover switch 2501 selects the vector output from the amplifier 2003 and selects this vector. Is output as a quantized prediction residual vector x _{n + 1} . In this case, it is not necessary to perform the process of generating vectors from the code book 2001 and the amplifiers 2002 and 2004 to the adder 2005.

The buffer 2502 holds the decoded LSF vector y _n of the previous frame for one frame, and outputs the decoded LSF vector y _n-1 of the previous frame to the amplifier 2004 and the amplifier 2503.

The amplifier 2004 multiplies the input decoded LSF vector y _n-1 two frames before by the weighting coefficient b ₋₁ and outputs the result to the adder 2005.

Adder 2005 calculates the sum of the vectors output from amplifier 2002 and amplifier 2004, and outputs the code vector that is the calculation result to adder 2504. That is, the adder 2005 calculates the vector x _n of the previous frame by weighted addition of the code vector specified by the LPC code L _{n + 1} of the current frame and the decoded LSF vector of the previous frame, and outputs it to the adder 2504 To do.

The amplifier 2503 multiplies the decoded LSF vector y _n-1 of the previous frame by the prediction coefficient a ₁ and outputs the result to the adder 2504.

The adder 2504 outputs the output of the adder 2005 (the decoded vector x _n of the previous frame recalculated using the LPC code L _{n + 1} of the current frame) and the output of the amplifier 2503 (the decoded LSF vector y _{n−1 of} the previous frame). (The vector multiplied by the prediction coefficient a ₁ ) and the decoded LSF vector y ′ _n of the previous frame is recalculated.

Note that the recalculation method of the decoded LSF vector y ′ _{n in} the present embodiment is the same as the compensation processing in the seventh embodiment.

As described above, according to the present embodiment, the decoding vector x _n obtained by the compensation processing of the seventh embodiment is used only for the predictor internal state at the time of decoding of the (n + 1) th frame, It is possible to reduce the processing delay of one frame required in the seventh embodiment.

(Embodiment 10)
In Embodiments 1 to 9 described above, only the configuration and processing in the LPC decoding unit are characterized. However, the configuration of the speech decoding apparatus according to the present embodiment is characterized in the configuration outside the LPC decoding unit. The present invention can be applied to any of FIGS. 1, 8, 11, and 21, but in this embodiment, a case where the present invention is applied to FIG. 21 will be described as an example.

  FIG. 26 is a block diagram showing the speech decoding apparatus according to the present embodiment. 26, the same components as those in FIG. 21 are denoted by the same reference numerals as those in FIG. 21, and detailed description thereof is omitted. The speech decoding apparatus 100 shown in FIG. 26 employs a configuration in which a filter gain calculation unit 2601, a sound source power control unit 2602, and an amplifier 2603 are added as compared to FIG. 21.

LPC decoding section 105 outputs the decoded LPC to LPC synthesis section 109 and filter gain calculation section 2601. In addition, LPC decoding section 105 outputs frame erasure code B _n corresponding to the n-th frame being decoded to excitation power control section 2602.

  The filter gain calculation unit 2601 calculates the filter gain of the synthesis filter constituted by the LPC input from the LPC decoding unit 105. As an example of the calculation method of the filter gain, there is a method of obtaining the square root of the energy of the impulse response to obtain the filter gain. This is based on the fact that if the input signal is considered as an impulse having an energy of 1, the energy of the impulse response of the synthesis filter constituted by the input LPC is directly used as filter gain information. Further, as another example of the filter gain calculation method, the mean square value of the linear prediction residual can be obtained from the LPC using the Levinson-Durbin algorithm. There is also a method of using the square root of the reciprocal of the mean square of the residual as the filter gain. The obtained filter gain is output to the sound source power control unit 2602. Note that the impulse response energy and the mean square value of the linear prediction residual may be output to the sound source power control unit 2602 without taking the square root as a parameter representing the filter gain.

The sound source power control unit 2602 receives the filter gain from the filter gain calculation unit 2601 and calculates a scaling coefficient for adjusting the amplitude of the sound source signal. The sound source power control unit 2602 includes a memory therein, and holds the filter gain of the previous frame in the memory. After the scaling coefficient is calculated, the contents of the memory are rewritten with the input filter gain of the current frame. The scaling factor SGn is calculated by assuming that the filter gain of the current frame is FG _n , the filter gain of the previous frame is FG _n−1 , and the upper limit value of the gain increase rate is DG _max , for example, SG _n = DG _max × FG _{n −1} / FG _n . Here, the gain increase rate is defined by FG _n / FG _n−1 and indicates how many times the filter gain of the current frame is larger than the filter gain of the previous frame. The upper limit value is determined in advance as DG _max . In the synthesis filter created by the frame erasure concealment process, when the filter gain increases rapidly with respect to the filter gain of the previous frame, the energy of the output signal of the synthesis filter also increases abruptly, and the decoded signal (synthesized signal) becomes local. Therefore, an abnormal noise is generated with a large amplitude. In order to avoid this, when the filter gain of the synthesis filter configured by the decoded LPC generated by the frame erasure concealment process becomes larger than a predetermined gain increase rate with respect to the filter gain of the previous frame, Reduce the power of the decoded excitation signal, which is the drive signal. A coefficient for this is a scaling coefficient, and the predetermined gain increase rate is an upper limit value DG _max of the gain increase rate. Usually, if DG _max is set to a value slightly smaller than 1, such as 1 or 0.98, the generation of abnormal noise can be avoided. Note that when FG _n / FG _n−1 is DG _max or less, SGn = 1.0 and scaling in the amplifier 2603 may not be performed.

Further, as another method for calculating the scaling coefficient SGn, for example, there is a method in which SGn = Max (SG _max , FG _n−1 / FG _n ) is obtained. Here, SG _max represents the maximum value of the scaling coefficient, and is a value slightly larger than 1, for example 1.5. Max (A, B) is a function that outputs the larger of A and B. In the case of SG _n = FG _n−1 / FG _n , the power of the sound source signal is reduced by the increase of the filter gain, and the energy of the decoded combined signal of the current frame becomes the same as the energy of the decoded combined signal of the previous frame. . As a result, it is possible to avoid the sudden increase of the combined signal energy described above and to avoid the sudden attenuation of the combined signal energy. If the filter gain of the current frame is smaller than the filter gain of the previous frame, the combined signal energy may be abruptly attenuated and perceived as sound interruption. In this _case, SG _n When _{SG n = FG n-1 /} FG n becomes a value of 1 or more, serves to avoid local attenuation of the composite signal energy. However, since the sound source signal generated by the frame erasure compensation process is not necessarily appropriate as the sound source signal, if the scaling coefficient is increased too much, the distortion becomes conspicuous and leads to quality degradation. For this reason, an upper limit is provided for the scaling coefficient, and when FG _n−1 / FG _n exceeds the upper limit value, clipping is performed to the upper limit value.

  It should be noted that the filter gain or the parameter representing the filter gain (such as the impulse response energy of the synthesis filter) one frame before is input from outside the sound source power control unit 2602 without being held in the memory in the sound source power control unit 2602. You may do it. In particular, when information relating to the filter gain of the previous frame is used in other parts of the speech decoder, the above parameters are input from the outside and are not rewritten inside the sound source power control unit 2602.

Then, the sound source power control unit 2602 receives the frame erasure code B _n from the LPC decoding unit 105, and when B _n indicates that the current frame is a erasure frame, the calculated scaling coefficient is supplied to the amplifier 2603. Output. On the other hand, when B _n indicates that the current frame is not an erasure frame, the sound source power control unit 2602 outputs 1 as a scaling coefficient to the amplifier 2603.

  The amplifier 2603 multiplies the decoded excitation signal input from the adder 108 by the scaling coefficient input from the excitation power control unit 2602 and outputs the result to the LPC synthesis unit 109.

  Thus, according to the present embodiment, when the filter gain of the synthesis filter configured by the decoded LPC generated by the frame erasure concealment process changes with respect to the filter gain of the previous frame, the synthesis filter drive signal By adjusting the power of the decoded audio signal, it is possible to prevent the generation of abnormal sounds and sound interruptions.

Note that even if B _n indicates that the current frame is not a lost frame, the previous frame is a lost frame (ie, B _n−1 indicates that the previous frame was a lost frame). The sound source power control unit 2602 may output the calculated scaling coefficient to the amplifier 2603. This is because, when predictive coding is used, the influence of errors may remain even in a return frame from frame loss. Even in this case, the same effect as described above can be obtained.

  The embodiment of the present invention has been described above.

  In each of the above embodiments, the encoding parameter is the LSF parameter. However, the present invention is not limited to this, and can be applied to any parameter as long as it varies gradually between frames. For example, immittance spectrum frequencies (ISFs) may be used.

  In each of the above-described embodiments, the encoding parameter is the LSF parameter itself. However, the difference from the average LSF may be taken, and the LSF parameter after removal of the average value may be used.

  The parameter decoding apparatus / parameter encoding apparatus according to the present invention is applicable to a speech decoding apparatus / speech encoding apparatus, and can be mounted on a communication terminal apparatus and a base station apparatus in a mobile communication system. Thus, it is possible to provide a communication terminal device, a base station device, and a mobile communication system having the same effects as described above.

  Further, here, the case where the present invention is configured by hardware has been described as an example, but the present invention can also be realized by software. For example, the algorithm of the parameter decoding method according to the present invention is described in a programming language, and this program is stored in a memory and executed by information processing means, thereby realizing the same function as the parameter decoding device according to the present invention. can do.

  Each functional block used in the description of each of the above embodiments is typically realized as an LSI which is an integrated circuit. These may be individually made into one chip, or may be made into one chip so as to include a part or all of them.

  Although referred to as LSI here, it may be called IC, system LSI, super LSI, ultra LSI, or the like depending on the degree of integration.

  Further, the method of circuit integration is not limited to LSI's, and implementation using dedicated circuitry or general purpose processors is also possible. An FPGA (Field Programmable Gate Array) that can be programmed after manufacturing the LSI or a reconfigurable processor that can reconfigure the connection or setting of circuit cells inside the LSI may be used.

  Further, if integrated circuit technology comes out to replace LSI's as a result of the advancement of semiconductor technology or a derivative other technology, it is naturally also possible to carry out function block integration using this technology. Biotechnology can be applied as a possibility.

  Japanese Patent Application No. 2006-305861 filed on Nov. 10, 2006, Japanese Patent Application No. 2007-132195 filed on May 17, 2007, and Japanese Patent Application No. 2007-240198 filed on Sep. 14, 2007, The entire disclosure of the drawings and abstract is incorporated herein by reference.

The parameter decoding apparatus and parameter decoding method according to the present invention can be applied to applications such as a speech decoding apparatus, a speech encoding apparatus, a communication terminal apparatus, a base station apparatus, and the like in a mobile communication system.

Claims (6)

A prediction residual decoding step for obtaining a quantized prediction residual of a spectrum parameter based on encoding information included in a current frame of a speech encoded bit string;
A parameter decoding step for decoding the spectral parameter based on the prediction residual;
Including
In the prediction residual decoding step, when the current frame is lost, the prediction residual of the current frame is obtained using the following equation:
Parameter decoding method.
x [n] = b0 x [n + 1] + b1 y [n-1]
where;
b0 = {b0 (i)}, i = 0,…, M-1,
b0 (i) = (1-a [1] '(i)) ([a [1]' (i)] ^ 2-2a [1] '(i) + 2) ^ (-1),
b1 = {b1 (i)}, i = 0,…, M-1, and
b1 (i) = ([a [1] '(i)] ^ 2-2a [1]' (i) + 2) ^ (-1)-a [1] (i).
where;
a [1] '(i): i-th component of the prediction coefficient set of the future frame,
a [1] (i): i-th component of the prediction coefficient set of the current frame,
x [n]: the prediction residual in the current frame,
x [n + 1]: the prediction residual in the future frame,
y [n-1]: Decoded spectrum parameter in the past frame The parameter decoding step includes:
A prediction step of multiplying a decoded spectrum parameter of a past frame by a prediction coefficient;
An addition step of decoding the spectral parameter by adding the prediction residual and the decoded spectral parameter of the past frame multiplied by the prediction coefficient;
The parameter decoding method according to claim 1, comprising:   The parameter decoding method according to claim 1, wherein the spectrum parameter is an ISF parameter. A prediction residual decoding means for obtaining a quantized prediction residual of a spectrum parameter based on coding information included in a current frame of a speech coded bit string;
Parameter decoding means for decoding the spectral parameter based on the prediction residual;
Including
In the prediction residual decoding means, when the current frame is lost, the prediction residual of the current frame is obtained using the following equation:
Parameter decoding device.
x [n] = b0 x [n + 1] + b1 y [n-1]
where;
b0 = {b0 (i)}, i = 0,…, M-1,
b0 (i) = (1-a [1] '(i)) ([a [1]' (i)] ^ 2-2a [1] '(i) + 2) ^ (-1),
b1 = {b1 (i)}, i = 0,…, M-1, and
b1 (i) = ([a [1] '(i)] ^ 2-2a [1]' (i) + 2) ^ (-1)-a [1] (i).
where;
a [1] '(i): i-th component of the prediction coefficient set of the future frame,
a [1] (i): i-th component of the prediction coefficient set of the current frame,
x [n]: the prediction residual in the current frame,
x [n + 1]: the prediction residual in the future frame,
y [n-1]: Decoded spectrum parameter in the past frame The parameter decoding means includes
Prediction means for multiplying the decoded spectrum parameter of the past frame by a prediction coefficient;
Adding means for decoding the spectral parameter by adding the prediction residual vector and the decoded spectral parameter of the past frame multiplied by the prediction coefficient;
The parameter decoding device according to claim 4, comprising:   The parameter decoding apparatus according to claim 4, wherein the spectrum parameter is an ISF parameter.

Images